Chunked Stencil Implementation PIM

libpimeval/src/libpimeval.cpp

@@ -501,17 +501,17 @@ pimRotateElementsLeft(PimObjId src)
 
 //! @brief  Shift elements of an obj by one step to the right and fill zero
 PimStatus
-pimShiftElementsRight(PimObjId src)
+pimShiftElementsRight(PimObjId src, bool useCrossRegionCommunication)
 {
-  bool ok = pimSim::get()->pimShiftElementsRight(src);
+  bool ok = pimSim::get()->pimShiftElementsRight(src, useCrossRegionCommunication);
   return ok ? PIM_OK : PIM_ERROR;
 }
 
 //! @brief  Shift elements of an obj by one step to the left and fill zero
 PimStatus
-pimShiftElementsLeft(PimObjId src)
+pimShiftElementsLeft(PimObjId src, bool useCrossRegionCommunication)
 {
-  bool ok = pimSim::get()->pimShiftElementsLeft(src);
+  bool ok = pimSim::get()->pimShiftElementsLeft(src, useCrossRegionCommunication);
   return ok ? PIM_OK : PIM_ERROR;
 }
 

libpimeval/src/libpimeval.h

@@ -193,8 +193,8 @@ PimStatus pimBroadcastUInt(PimObjId dest, uint64_t value);
 PimStatus pimBroadcastFP(PimObjId dest, float value);
 PimStatus pimRotateElementsRight(PimObjId src);
 PimStatus pimRotateElementsLeft(PimObjId src);
-PimStatus pimShiftElementsRight(PimObjId src);
-PimStatus pimShiftElementsLeft(PimObjId src);
+PimStatus pimShiftElementsRight(PimObjId src, bool useCrossRegionCommunication = true);
+PimStatus pimShiftElementsLeft(PimObjId src, bool useCrossRegionCommunication = true);
 PimStatus pimShiftBitsRight(PimObjId src, PimObjId dest, unsigned shiftAmount);
 PimStatus pimShiftBitsLeft(PimObjId src, PimObjId dest, unsigned shiftAmount);
 

libpimeval/src/pimCmd.cpp

@@ -1205,33 +1205,35 @@ pimCmdRotate::execute()
   computeAllRegions(numRegions);
 
   // handle region boundaries
-  if (m_cmdType == PimCmdEnum::ROTATE_ELEM_R || m_cmdType == PimCmdEnum::SHIFT_ELEM_R) {
-    for (unsigned i = 0; i < numRegions; ++i) {
-      const pimRegion &srcRegion = objSrc.getRegions()[i];
-      uint64_t elemIdxBegin = srcRegion.getElemIdxBegin();
-      uint64_t val = 0;
-      if (i == 0 && m_cmdType == PimCmdEnum::ROTATE_ELEM_R) {
-        val = m_regionBoundary[numRegions - 1];
-      } else if (i > 0) {
-        val = m_regionBoundary[i - 1];
+  if(m_useCrossRegionCommunication) {
+    if (m_cmdType == PimCmdEnum::ROTATE_ELEM_R || m_cmdType == PimCmdEnum::SHIFT_ELEM_R) {
+      for (unsigned i = 0; i < numRegions; ++i) {
+        const pimRegion &srcRegion = objSrc.getRegions()[i];
+        uint64_t elemIdxBegin = srcRegion.getElemIdxBegin();
+        uint64_t val = 0;
+        if (i == 0 && m_cmdType == PimCmdEnum::ROTATE_ELEM_R) {
+          val = m_regionBoundary[numRegions - 1];
+        } else if (i > 0) {
+          val = m_regionBoundary[i - 1];
+        }
+        objSrc.setElement(elemIdxBegin, val);
       }
-      objSrc.setElement(elemIdxBegin, val);
-    }
-  } else if (m_cmdType == PimCmdEnum::ROTATE_ELEM_L || m_cmdType == PimCmdEnum::SHIFT_ELEM_L) {
-    for (unsigned i = 0; i < numRegions; ++i) {
-      const pimRegion &srcRegion = objSrc.getRegions()[i];
-      unsigned numElementsInRegion = srcRegion.getNumElemInRegion();
-      uint64_t elemIdxBegin = srcRegion.getElemIdxBegin();
-      uint64_t val = 0;
-      if (i == numRegions - 1 && m_cmdType == PimCmdEnum::ROTATE_ELEM_L) {
-        val = m_regionBoundary[0];
-      } else if (i < numRegions - 1) {
-        val = m_regionBoundary[i + 1];
+    } else if (m_cmdType == PimCmdEnum::ROTATE_ELEM_L || m_cmdType == PimCmdEnum::SHIFT_ELEM_L) {
+      for (unsigned i = 0; i < numRegions; ++i) {
+        const pimRegion &srcRegion = objSrc.getRegions()[i];
+        unsigned numElementsInRegion = srcRegion.getNumElemInRegion();
+        uint64_t elemIdxBegin = srcRegion.getElemIdxBegin();
+        uint64_t val = 0;
+        if (i == numRegions - 1 && m_cmdType == PimCmdEnum::ROTATE_ELEM_L) {
+          val = m_regionBoundary[0];
+        } else if (i < numRegions - 1) {
+          val = m_regionBoundary[i + 1];
+        }
+        objSrc.setElement(elemIdxBegin + numElementsInRegion - 1, val);
       }
-      objSrc.setElement(elemIdxBegin + numElementsInRegion - 1, val);
+    } else {
+      assert(0);
     }
-  } else {
-    assert(0);
   }
 
   if (pimSim::get()->getDeviceType() != PIM_FUNCTIONAL) {
@@ -1306,7 +1308,7 @@ pimCmdRotate::updateStats() const
   PimDataType dataType = objSrc.getDataType();
   bool isVLayout = objSrc.isVLayout();
 
-  pimeval::perfEnergy mPerfEnergy = pimSim::get()->getPerfEnergyModel()->getPerfEnergyForRotate(m_cmdType, objSrc);
+  pimeval::perfEnergy mPerfEnergy = pimSim::get()->getPerfEnergyModel()->getPerfEnergyForRotate(m_cmdType, objSrc, m_useCrossRegionCommunication);
   pimSim::get()->getStatsMgr()->recordCmd(getName(dataType, isVLayout), mPerfEnergy);
   return true;
 }

libpimeval/src/pimCmd.h

@@ -506,8 +506,8 @@ class pimCmdBroadcast : public pimCmd
 class pimCmdRotate : public pimCmd
 {
 public:
-  pimCmdRotate(PimCmdEnum cmdType, PimObjId src)
-    : pimCmd(cmdType), m_src(src)
+  pimCmdRotate(PimCmdEnum cmdType, PimObjId src, bool useCrossRegionCommunication)
+    : pimCmd(cmdType), m_src(src), m_useCrossRegionCommunication(useCrossRegionCommunication)
   {
     assert(cmdType == PimCmdEnum::ROTATE_ELEM_R || cmdType == PimCmdEnum::ROTATE_ELEM_L ||
            cmdType == PimCmdEnum::SHIFT_ELEM_R || cmdType == PimCmdEnum::SHIFT_ELEM_L);
@@ -520,6 +520,7 @@ class pimCmdRotate : public pimCmd
 protected:
   PimObjId m_src;
   std::vector<uint64_t> m_regionBoundary;
+  bool m_useCrossRegionCommunication;
 };
 
 //! @class  pimCmdReadRowToSa

libpimeval/src/pimPerfEnergyAquabolt.cpp

@@ -237,7 +237,7 @@ pimPerfEnergyAquabolt::getPerfEnergyForBroadcast(PimCmdEnum cmdType, const pimOb
 
 //! @brief  Perf energy model of aquabolt PIM for rotate
 pimeval::perfEnergy
-pimPerfEnergyAquabolt::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj) const
+pimPerfEnergyAquabolt::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj, bool useCrossRegionCommunication) const
 {
   double msRuntime = 0.0;
   double mjEnergy = 0.0;

libpimeval/src/pimPerfEnergyAquabolt.h

@@ -26,7 +26,7 @@ class pimPerfEnergyAquabolt : public pimPerfEnergyBase
   virtual pimeval::perfEnergy getPerfEnergyForFunc2(PimCmdEnum cmdType, const pimObjInfo& objSrc1, const pimObjInfo& objSrc2, const pimObjInfo& objDest) const override;
   virtual pimeval::perfEnergy getPerfEnergyForReduction(PimCmdEnum cmdType, const pimObjInfo& obj, unsigned numPass) const override;
   virtual pimeval::perfEnergy getPerfEnergyForBroadcast(PimCmdEnum cmdType, const pimObjInfo& obj) const override;
-  virtual pimeval::perfEnergy getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj) const override;
+  virtual pimeval::perfEnergy getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj, bool useCrossRegionCommunication) const override;
 
 protected:
   unsigned m_aquaboltFPUBitWidth = 16;

libpimeval/src/pimPerfEnergyBankLevel.cpp

@@ -297,7 +297,7 @@ pimPerfEnergyBankLevel::getPerfEnergyForBroadcast(PimCmdEnum cmdType, const pimO
 // TODO: This needs to be revisited
 //! @brief  Perf energy model of bank-level PIM for rotate
 pimeval::perfEnergy
-pimPerfEnergyBankLevel::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj) const
+pimPerfEnergyBankLevel::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj, bool useCrossRegionCommunication) const
 {
   double msRuntime = 0.0;
   double mjEnergy = 0.0;
@@ -308,8 +308,6 @@ pimPerfEnergyBankLevel::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjI
   unsigned bitsPerElement = obj.getBitsPerElement(PimBitWidth::ACTUAL);
   unsigned numRegions = obj.getRegions().size();
   uint64_t totalOp = 0;
-  // boundary handling - assume two times copying between device and host for boundary elements
-  pimeval::perfEnergy perfEnergyBT = getPerfEnergyForBytesTransfer(PimCmdEnum::COPY_D2H, numRegions * bitsPerElement / 8);
 
   // rotate within subarray:
   // For every bit: Read row to SA; move SA to R1; Shift R1 by N steps; Move R1 to SA; Write SA to row
@@ -319,8 +317,12 @@ pimPerfEnergyBankLevel::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjI
   msRuntime = (m_tR + (bitsPerElement + 2) * m_tL + m_tW); // for one pass
   msRuntime *= numPass;
   mjEnergy = (m_eAP + (bitsPerElement + 2) * m_eL) * numPass;
-  msRuntime += 2 * perfEnergyBT.m_msRuntime;
-  mjEnergy += 2 * perfEnergyBT.m_mjEnergy;
+  if(useCrossRegionCommunication) {
+    // boundary handling - assume two times copying between device and host for boundary elements
+    pimeval::perfEnergy perfEnergyBT = getPerfEnergyForBytesTransfer(PimCmdEnum::COPY_D2H, numRegions * bitsPerElement / 8);
+    msRuntime += 2 * perfEnergyBT.m_msRuntime;
+    mjEnergy += 2 * perfEnergyBT.m_mjEnergy;
+  }
   std::cout << "PIM-Warning: Perf energy model is not precise for PIM command " << pimCmd::getName(cmdType, "") << std::endl;
 
   return pimeval::perfEnergy(msRuntime, mjEnergy, msRead, msWrite, msCompute, totalOp);

libpimeval/src/pimPerfEnergyBankLevel.h

@@ -26,7 +26,7 @@ class pimPerfEnergyBankLevel : public pimPerfEnergyBase
   virtual pimeval::perfEnergy getPerfEnergyForFunc2(PimCmdEnum cmdType, const pimObjInfo& objSrc1, const pimObjInfo& objSrc2, const pimObjInfo& objDest) const override;
   virtual pimeval::perfEnergy getPerfEnergyForReduction(PimCmdEnum cmdType, const pimObjInfo& obj, unsigned numPass) const override;
   virtual pimeval::perfEnergy getPerfEnergyForBroadcast(PimCmdEnum cmdType, const pimObjInfo& obj) const override;
-  virtual pimeval::perfEnergy getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj) const override;
+  virtual pimeval::perfEnergy getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj, bool useCrossRegionCommunication) const override;
 
 protected:
   double m_blimpCoreOriginalLatency = 0.000005; // ms; 200 MHz. Reference: BLIMP paper

libpimeval/src/pimPerfEnergyBase.cpp

@@ -156,7 +156,7 @@ pimPerfEnergyBase::getPerfEnergyForBroadcast(PimCmdEnum cmdType, const pimObjInf
 
 //! @brief  Perf energy model of base class for rotate (placeholder)
 pimeval::perfEnergy
-pimPerfEnergyBase::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj) const
+pimPerfEnergyBase::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj, bool useCrossRegionCommunication) const
 {
   double msRuntime = 1e10;
   double mjEnergy = 999999999.9;

libpimeval/src/pimPerfEnergyBase.h

@@ -70,7 +70,7 @@ class pimPerfEnergyBase
   virtual pimeval::perfEnergy getPerfEnergyForFunc2(PimCmdEnum cmdType, const pimObjInfo& objSrc1, const pimObjInfo& objSrc2, const pimObjInfo& objDest) const;
   virtual pimeval::perfEnergy getPerfEnergyForReduction(PimCmdEnum cmdType, const pimObjInfo& obj, unsigned numPass) const;
   virtual pimeval::perfEnergy getPerfEnergyForBroadcast(PimCmdEnum cmdType, const pimObjInfo& obj) const;
-  virtual pimeval::perfEnergy getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj) const;
+  virtual pimeval::perfEnergy getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj, bool useCrossRegionCommunication) const;
 
 protected:
   PimDeviceEnum m_simTarget;

libpimeval/src/pimPerfEnergyBitSerial.cpp

@@ -441,7 +441,7 @@ pimPerfEnergyBitSerial::getPerfEnergyForBroadcast(PimCmdEnum cmdType, const pimO
 
 //! @brief  Perf energy model of bit-serial PIM for rotate
 pimeval::perfEnergy
-pimPerfEnergyBitSerial::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj) const
+pimPerfEnergyBitSerial::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj, bool useCrossRegionCommunication) const
 {
   double msRuntime = 0.0;
   double mjEnergy = 0.0;
@@ -453,8 +453,6 @@ pimPerfEnergyBitSerial::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjI
   unsigned bitsPerElement = obj.getBitsPerElement(PimBitWidth::ACTUAL);
   unsigned numRegions = obj.getRegions().size();
   unsigned numCore = obj.getNumCoreAvailable();
-  // boundary handling - assume two times copying between device and host for boundary elements
-  pimeval::perfEnergy perfEnergyBT = getPerfEnergyForBytesTransfer(PimCmdEnum::COPY_D2H, numRegions * bitsPerElement / 8);
 
   switch (m_simTarget) {
     case PIM_DEVICE_BITSIMD_V:
@@ -467,8 +465,12 @@ pimPerfEnergyBitSerial::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjI
       totalOp += 3 * bitsPerElement * numPass * numCore;
       msRuntime = msRead + msWrite + msCompute;
       mjEnergy = (m_eAP + 3 * m_eL) * bitsPerElement * numPass; // for one pass
-      msRuntime += 2 * perfEnergyBT.m_msRuntime;
-      mjEnergy += 2 * perfEnergyBT.m_mjEnergy;
+      if(useCrossRegionCommunication) {
+        // boundary handling - assume two times copying between device and host for boundary elements
+        pimeval::perfEnergy perfEnergyBT = getPerfEnergyForBytesTransfer(PimCmdEnum::COPY_D2H, numRegions * bitsPerElement / 8);
+        msRuntime += 2 * perfEnergyBT.m_msRuntime;
+        mjEnergy += 2 * perfEnergyBT.m_mjEnergy;
+      }
       break;
     case PIM_DEVICE_SIMDRAM:
       // todo
@@ -483,8 +485,12 @@ pimPerfEnergyBitSerial::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjI
       msRuntime = (m_tR + (bitsPerElement + 2) * m_tL + m_tW); // for one pass
       msRuntime *= numPass;
       mjEnergy = (m_eAP + (bitsPerElement + 2) * m_eL) * numPass;
-      msRuntime += 2 * perfEnergyBT.m_msRuntime;
-      mjEnergy += 2 * perfEnergyBT.m_mjEnergy;
+      if(useCrossRegionCommunication) {
+        // boundary handling - assume two times copying between device and host for boundary elements
+        pimeval::perfEnergy perfEnergyBT = getPerfEnergyForBytesTransfer(PimCmdEnum::COPY_D2H, numRegions * bitsPerElement / 8);
+        msRuntime += 2 * perfEnergyBT.m_msRuntime;
+        mjEnergy += 2 * perfEnergyBT.m_mjEnergy;
+      }
       break;
     default:
       assert(0);

libpimeval/src/pimPerfEnergyBitSerial.h

@@ -26,7 +26,7 @@ class pimPerfEnergyBitSerial : public pimPerfEnergyBase
   virtual pimeval::perfEnergy getPerfEnergyForFunc2(PimCmdEnum cmdType, const pimObjInfo& objSrc1, const pimObjInfo& objSrc2, const pimObjInfo& objDest) const override;
   virtual pimeval::perfEnergy getPerfEnergyForReduction(PimCmdEnum cmdType, const pimObjInfo& obj, unsigned numPass) const override;
   virtual pimeval::perfEnergy getPerfEnergyForBroadcast(PimCmdEnum cmdType, const pimObjInfo& obj) const override;
-  virtual pimeval::perfEnergy getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj) const override;
+  virtual pimeval::perfEnergy getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj, bool useCrossRegionCommunication) const override;
 
 protected:
   pimeval::perfEnergy getPerfEnergyBitSerial(PimDeviceEnum deviceType, PimCmdEnum cmdType, unsigned numPass, const pimObjInfo& objSrc1, const pimObjInfo& objSrc2, const pimObjInfo& objDest) const;

libpimeval/src/pimPerfEnergyFulcrum.cpp

@@ -301,7 +301,7 @@ pimPerfEnergyFulcrum::getPerfEnergyForBroadcast(PimCmdEnum cmdType, const pimObj
 
 //! @brief  Perf energy model of Fulcrum for rotate
 pimeval::perfEnergy
-pimPerfEnergyFulcrum::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj) const
+pimPerfEnergyFulcrum::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj, bool useCrossRegionCommunication) const
 {
   double msRuntime = 0.0;
   double mjEnergy = 0.0;
@@ -312,8 +312,6 @@ pimPerfEnergyFulcrum::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInf
   unsigned bitsPerElement = obj.getBitsPerElement(PimBitWidth::ACTUAL);
   unsigned numRegions = obj.getRegions().size();
   uint64_t totalOp = 0;
-  // boundary handling - assume two times copying between device and host for boundary elements
-  pimeval::perfEnergy perfEnergyBT = getPerfEnergyForBytesTransfer(PimCmdEnum::COPY_D2H, numRegions * bitsPerElement / 8);
 
   // rotate within subarray:
   // For every bit: Read row to SA; move SA to R1; Shift R1 by N steps; Move R1 to SA; Write SA to row
@@ -324,8 +322,12 @@ pimPerfEnergyFulcrum::getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInf
   msWrite = m_tW * numPass;
   msRuntime = msRead + msWrite + msCompute;
   mjEnergy = (m_eAP + (bitsPerElement + 2) * m_eL) * numPass;
-  msRuntime += 2 * perfEnergyBT.m_msRuntime;
-  mjEnergy += 2 * perfEnergyBT.m_mjEnergy;
+  if(useCrossRegionCommunication) {
+    // boundary handling - assume two times copying between device and host for boundary elements
+    pimeval::perfEnergy perfEnergyBT = getPerfEnergyForBytesTransfer(PimCmdEnum::COPY_D2H, numRegions * bitsPerElement / 8);
+    msRuntime += 2 * perfEnergyBT.m_msRuntime;
+    mjEnergy += 2 * perfEnergyBT.m_mjEnergy;
+  }
   std::cout << "PIM-Warning: Perf energy model is not precise for PIM command " << pimCmd::getName(cmdType, "") << std::endl;
 
   return pimeval::perfEnergy(msRuntime, mjEnergy, msRead, msWrite, msCompute, totalOp);

libpimeval/src/pimPerfEnergyFulcrum.h

@@ -26,7 +26,7 @@ class pimPerfEnergyFulcrum : public pimPerfEnergyBase
   virtual pimeval::perfEnergy getPerfEnergyForFunc2(PimCmdEnum cmdType, const pimObjInfo& objSrc1, const pimObjInfo& objSrc2, const pimObjInfo& objDest) const override;
   virtual pimeval::perfEnergy getPerfEnergyForReduction(PimCmdEnum cmdType, const pimObjInfo& obj, unsigned numPass) const override;
   virtual pimeval::perfEnergy getPerfEnergyForBroadcast(PimCmdEnum cmdType, const pimObjInfo& obj) const override;
-  virtual pimeval::perfEnergy getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj) const override;
+  virtual pimeval::perfEnergy getPerfEnergyForRotate(PimCmdEnum cmdType, const pimObjInfo& obj, bool useCrossRegionCommunication) const override;
 
 protected:
   double m_fulcrumMulLatency = 0.00000609; // 6.09ns

libpimeval/src/pimSim.cpp

@@ -826,7 +826,7 @@ pimSim::pimRotateElementsRight(PimObjId src)
 {
   pimPerfMon perfMon("pimRotateElementsRight");
   if (!isValidDevice()) { return false; }
-  std::unique_ptr<pimCmd> cmd = std::make_unique<pimCmdRotate>(PimCmdEnum::ROTATE_ELEM_R, src);
+  std::unique_ptr<pimCmd> cmd = std::make_unique<pimCmdRotate>(PimCmdEnum::ROTATE_ELEM_R, src, true);
   return m_device->executeCmd(std::move(cmd));
 }
 
@@ -835,25 +835,25 @@ pimSim::pimRotateElementsLeft(PimObjId src)
 {
   pimPerfMon perfMon("pimRotateElementsLeft");
   if (!isValidDevice()) { return false; }
-  std::unique_ptr<pimCmd> cmd = std::make_unique<pimCmdRotate>(PimCmdEnum::ROTATE_ELEM_L, src);
+  std::unique_ptr<pimCmd> cmd = std::make_unique<pimCmdRotate>(PimCmdEnum::ROTATE_ELEM_L, src, true);
   return m_device->executeCmd(std::move(cmd));
 }
 
 bool
-pimSim::pimShiftElementsRight(PimObjId src)
+pimSim::pimShiftElementsRight(PimObjId src, bool useCrossRegionCommunication)
 {
   pimPerfMon perfMon("pimShiftElementsRight");
   if (!isValidDevice()) { return false; }
-  std::unique_ptr<pimCmd> cmd = std::make_unique<pimCmdRotate>(PimCmdEnum::SHIFT_ELEM_R, src);
+  std::unique_ptr<pimCmd> cmd = std::make_unique<pimCmdRotate>(PimCmdEnum::SHIFT_ELEM_R, src, useCrossRegionCommunication);
   return m_device->executeCmd(std::move(cmd));
 }
 
 bool
-pimSim::pimShiftElementsLeft(PimObjId src)
+pimSim::pimShiftElementsLeft(PimObjId src, bool useCrossRegionCommunication)
 {
   pimPerfMon perfMon("pimShiftElementsLeft");
   if (!isValidDevice()) { return false; }
-  std::unique_ptr<pimCmd> cmd = std::make_unique<pimCmdRotate>(PimCmdEnum::SHIFT_ELEM_L, src);
+  std::unique_ptr<pimCmd> cmd = std::make_unique<pimCmdRotate>(PimCmdEnum::SHIFT_ELEM_L, src, useCrossRegionCommunication);
   return m_device->executeCmd(std::move(cmd));
 }
 

libpimeval/src/pimSim.h

@@ -120,8 +120,8 @@ class pimSim
   template <typename T> bool pimBroadcast(PimObjId dest, T value);
   bool pimRotateElementsRight(PimObjId src);
   bool pimRotateElementsLeft(PimObjId src);
-  bool pimShiftElementsRight(PimObjId src);
-  bool pimShiftElementsLeft(PimObjId src);
+  bool pimShiftElementsRight(PimObjId src, bool useCrossRegionCommunication);
+  bool pimShiftElementsLeft(PimObjId src, bool useCrossRegionCommunication);
   bool pimShiftBitsRight(PimObjId src, PimObjId dest, unsigned shiftAmount);
   bool pimShiftBitsLeft(PimObjId src, PimObjId dest, unsigned shiftAmount);
 

misc-bench/stencil/Makefile

@@ -0,0 +1,16 @@
+# Makefile: C++ version of stencil
+# Copyright (c) 2025 University of Virginia
+# This file is licensed under the MIT License.
+# See the LICENSE file in the root of this repository for more details.
+
+SUBDIRS := PIM
+
+.PHONY: debug perf dramsim3_integ clean $(SUBDIRS)
+.DEFAULT_GOAL := perf
+
+USE_OPENMP ?= 0
+
+debug perf dramsim3_integ clean: $(SUBDIRS)
+
+$(SUBDIRS):
+	$(MAKE) -C $@ $(MAKECMDGOALS) USE_OPENMP=$(USE_OPENMP)