Adding asynchronous temporal partitioning

Author: marjanfariborz

src/accl/graph/sega/MPU.py

@@ -27,9 +27,10 @@
 
 from m5.params import *
 from m5.proxy import *
-from m5.SimObject import SimObject
+# from m5.SimObject import SimObject
+from m5.objects.ClockedObject import ClockedObject
 
-class MPU(SimObject):
+class MPU(ClockedObject):
     type = "MPU"
     cxx_header = "accl/graph/sega/mpu.hh"
     cxx_class = "gem5::MPU"

src/accl/graph/sega/centeral_controller.cc

@@ -93,6 +93,18 @@ CenteralController::createBCWorkload(Addr init_addr, uint32_t init_value)
     workload = new BSPBCWorkload(init_addr, init_value);
 }
 
+bool
+CenteralController::bufferRemoteUpdate(int slice_number, PacketPtr pkt)
+{
+    for (auto mpu: mpuVector) {
+        if (contains(mpu->getAddrRanges(), pkt->getAddr())) {
+            remoteUpdates[mpu][slice_number].push_back(pkt);
+        }
+    }
+    
+    return true;
+}
+
 void
 CenteralController::createPopCountDirectory(int atoms_per_block)
 {
@@ -173,6 +185,25 @@ CenteralController::recvDoneSignal()
     bool done = true;
     for (auto mpu : mpuVector) {
         done &= mpu->done();
+        int total_num_slices = remoteUpdates[mpu].size();
+        if (mpu->done()) {
+            int slice_number = mpu->getSliceCounter() + 1;
+            while ((total_num_slices != 0) && (slice_number != mpu->getSliceCounter())) {
+                if (!remoteUpdates[mpu][slice_number].empty()) {
+                    mpu->scheduleNewSlice();
+                    mpu->updateSliceCounter(slice_number);
+                    done = false;
+                    break;
+                } 
+                else {
+                    if (slice_number == total_num_slices) {
+                        slice_number = 0;
+                    } else {
+                        slice_number++;
+                    }
+                }
+            }
+        }
     }
 
     if (done && mode == ProcessingMode::ASYNCHRONOUS) {

src/accl/graph/sega/centeral_controller.hh

@@ -77,11 +77,16 @@ class CenteralController : public ClockedObject
     void createPRWorkload(int num_nodes, float alpha);
     void createBCWorkload(Addr init_addr, uint32_t init_value);
 
+    bool bufferRemoteUpdate(int slice_number, PacketPtr pkt);
+    int getnumGPTs() {return mpuVector.size();}
+
     void recvDoneSignal();
 
     int workCount();
     float getPRError();
     void printAnswerToHostSimout();
+    std::unordered_map<MPU*, std::unordered_map<int, std::deque<PacketPtr>>> 
+                                                                remoteUpdates; 
 };
 
 }

src/accl/graph/sega/coalesce_engine.hh

@@ -217,6 +217,10 @@ class CoalesceEngine : public BaseMemoryEngine
     ReadReturnStatus recvWLRead(Addr addr);
     void recvWLWrite(Addr addr, WorkListItem wl);
 
+    int getSliceSize() 
+                    {return (int)(params().cache_size); }
+                    // /sizeof(WorkListItem)); }
+
     int workCount();
     int futureWorkCount();
     void recvVertexPull();