Merge branch 'devel/fix_cuda_nodePos_comm' into 'master'

devel/fix_cuda_nodePos_comm

See merge request exastencils/exastencils!72

.gitlab-ci.yml

@@ -52,6 +52,7 @@ stages:
         - python3 --version
         - mpirun --version
         - nvcc --version
+        - nvidia-smi
         - updatedb
         - locate cuda.h
         - locate libcudart

Compiler/src/exastencils/app/ir/IR_LayerHandler.scala

@@ -325,10 +325,18 @@ object IR_DefaultLayerHandler extends IR_LayerHandler {
     if (Knowledge.data_genVariableFieldSizes)
       IR_GenerateIndexManipFcts.apply()
 
+    // adapt accesses to device data in case of managed memory
+    if (Knowledge.cuda_enabled && Knowledge.cuda_useManagedMemory)
+      CUDA_AdaptDeviceAccessesForMM.apply()
+
     IR_AddInternalVariables.apply()
     // resolve possibly newly added constant IVs
     IR_ResolveConstIVs.apply()
 
+    // adapt allocations and de-allocations before expanding
+    if (Knowledge.cuda_enabled)
+      CUDA_AdaptAllocations.apply()
+
     if (Knowledge.useFasterExpand)
       IR_ExpandInOnePass.apply()
     else

Compiler/src/exastencils/applications/ir/IR_HandleMainApplication.scala

@@ -26,7 +26,6 @@ import exastencils.base.ir._
 import exastencils.config.Knowledge
 import exastencils.datastructures._
 import exastencils.logger.Logger
-import exastencils.parallelization.api.cuda._
 import exastencils.parallelization.api.mpi._
 import exastencils.parallelization.api.omp.OMP_Parallel
 
@@ -51,11 +50,6 @@ object IR_HandleMainApplication extends DefaultStrategy("HandleMainApplication")
         func.body.append(IR_Native("LIKWID_MARKER_CLOSE"))
       }
 
-      if (Knowledge.cuda_enabled) {
-        func.body.prepend(CUDA_Init)
-        func.body.append(CUDA_Finalize)
-      }
-
       if (Knowledge.mpi_enabled) {
         func.body.prepend(MPI_Init)
         func.body.append(MPI_Finalize)

Compiler/src/exastencils/base/ir/IR_Alloc.scala

@@ -38,7 +38,7 @@ case class IR_ArrayAllocation(var name : IR_Expression, // no string - could be
 
 case class IR_ScalarAllocation(var name : IR_Expression, // no string - could be an IV
     var datatype : IR_Datatype,
-    ) extends IR_Statement {
+) extends IR_Statement {
   //override def datatype = IR_UnitDatatype
   override def prettyprint(out : PpStream) : Unit = out << name << " = " << "new" << ' ' << datatype << ";"
 }
@@ -49,7 +49,6 @@ case class IR_ArrayFree(var pointer : IR_Expression) extends IR_Statement {
   override def prettyprint(out : PpStream) : Unit = out << "delete[] " << pointer << ";"
 }
 
-
 /// IR_ScalarFree
 case class IR_ScalarFree(var pointer : IR_Expression) extends IR_Statement {
   override def prettyprint(out : PpStream) : Unit = out << "delete " << pointer << ";"

Compiler/src/exastencils/config/Knowledge.scala

@@ -515,6 +515,16 @@ object Knowledge {
   // apply spatial blocking with read-only cache
   var cuda_spatialBlockingWithROC : Boolean = false
 
+  // use pinned memory to allocate host field data and buffers
+  var cuda_usePinnedHostMemory : Boolean = true
+  // use managed memory instead of host and device variants for field data and buffers
+  var cuda_useManagedMemory : Boolean = false
+  // replace device variants of field data and buffers with device pointers derived from host counter-parts
+  var cuda_useZeroCopy : Boolean = false
+
+  // only relevant if cuda_useManagedMemory == true; replace cuda memcpy with asynchronous prefetches
+  var cuda_genAsyncPrefetch : Boolean = true
+
   // if true, the first dimension of the block size is enlarged if the kernel dimensionality is lower than the global dimensionality
   var cuda_foldBlockSizeForRedDimensionality : Boolean = true
 
@@ -647,9 +657,6 @@ object Knowledge {
   var experimental_CTPivotElimination : Boolean = false
   var experimental_QRPivot : Boolean = false
 
-  // eliminate occurrences of cudaContext - required for PizDaint
-  var experimental_eliminateCudaContext : Boolean = false
-
   // generate cuda kernels independently of them being parallel or not
   var experimental_cuda_generateKernelForNonParallel : Boolean = false
 
@@ -672,7 +679,6 @@ object Knowledge {
   var visit_enable : Boolean = false
   var experimental_visit_addCurveMesh : Boolean = false
 
-
   /// === constraints and resolutions ===
   def update() : Unit = {
     // NOTE: it is required to call update at least once
@@ -774,7 +780,13 @@ object Knowledge {
     Constraints.condWarn(cuda_enabled && opt_conventionalCSE && !useDblPrecision, "Double precision should be used if CUDA is enabled and CSE should be applied!")
     Constraints.condEnsureValue(useDblPrecision, true, cuda_enabled && opt_conventionalCSE)
 
-    Constraints.condError(!cuda_memory_transfer_elimination_options.contains(cuda_eliminate_memory_transfers), "Invalid value for \"cuda_eliminate_memory_transfers\". Should be one of: " + cuda_memory_transfer_elimination_options.mkString(",") )
+    Constraints.condError(!cuda_memory_transfer_elimination_options.contains(cuda_eliminate_memory_transfers), "Invalid value for \"cuda_eliminate_memory_transfers\". Should be one of: " + cuda_memory_transfer_elimination_options.mkString(","))
+
+    Constraints.condEnsureValue(cuda_usePinnedHostMemory, true, cuda_useZeroCopy)
+    Constraints.condError(cuda_useManagedMemory && cuda_usePinnedHostMemory, "cuda_useManagedMemory and cuda_usePinnedHostMemory are mutually exclusive")
+    Constraints.condError(cuda_useManagedMemory && cuda_useZeroCopy, "cuda_useManagedMemory and cuda_usePinnedHostMemory are mutually exclusive")
+    Constraints.condEnsureValue(data_alignFieldPointers, false, cuda_useManagedMemory || cuda_usePinnedHostMemory)
+    Constraints.condEnsureValue(data_alignTmpBufferPointers, false, cuda_useManagedMemory || cuda_usePinnedHostMemory)
 
     Constraints.condWarn(experimental_splitLoopsForAsyncComm && !comm_onlyAxisNeighbors, s"Using asynchronous communication with comm_onlyAxisNeighbors leads to problems with stencils containing diagonal entries")
 

Compiler/src/exastencils/globals/ir/IR_AddInternalVariables.scala

@@ -203,7 +203,7 @@ object IR_AddInternalVariables extends DefaultStrategy("Add internal variables")
         bufferAllocs += (id -> IR_LoopOverFragments(
           IR_ArrayAllocation(
             buf,
-            if(buf.field.layout.datatype.isInstanceOf[IR_ComplexDatatype]) buf.field.layout.datatype
+            if (buf.field.layout.datatype.isInstanceOf[IR_ComplexDatatype]) buf.field.layout.datatype
             else IR_RealDatatype,
             size
           ), IR_ParallelizationInfo(potentiallyParallel = true)))
@@ -259,9 +259,11 @@ object IR_AddInternalVariables extends DefaultStrategy("Add internal variables")
           func.body += genericAlloc._2
 
       for (deviceAlloc <- deviceFieldAllocs.toSeq.sortBy(_._1) ++ deviceBufferAllocs.toSeq.sortBy(_._1))
-        if ("Condition" == Knowledge.cuda_preferredExecution)
-          func.body += IR_IfCondition(IR_Negation(Knowledge.cuda_executionCondition), deviceAlloc._2)
-        else if ("MSVC" == Platform.targetCompiler /*&& Platform.targetCompilerVersion <= 11*/ ) // fix for https://support.microsoft.com/en-us/kb/315481
+        if ("Condition" == Knowledge.cuda_preferredExecution) {
+          val loop = deviceAlloc._2.asInstanceOf[IR_LoopOverFragments]
+          loop.body = ListBuffer(IR_IfCondition(IR_Negation(Knowledge.cuda_executionCondition), loop.body))
+          func.body += loop
+        } else if ("MSVC" == Platform.targetCompiler /*&& Platform.targetCompilerVersion <= 11*/ ) // fix for https://support.microsoft.com/en-us/kb/315481
           func.body += IR_Scope(deviceAlloc._2)
         else
           func.body += deviceAlloc._2

Compiler/src/exastencils/optimization/ir/IR_Vectorization.scala

@@ -79,6 +79,7 @@ private object VectorizeInnermost extends PartialFunction[Node, Transformation.O
       case ex : VectorizationException =>
         if (DEBUG) {
           val msg : String = "[vect]  unable to vectorize loop: " + ex.msg + "  (line " + ex.getStackTrace()(0).getLineNumber + ')'
+          Logger.warn(msg)
           println(msg) // print directly, logger may be silenced by any surrounding strategy
           return List(IR_Comment(msg), node)
         }
@@ -144,6 +145,8 @@ private object VectorizeInnermost extends PartialFunction[Node, Transformation.O
     private var alignedResidue : Long = -1
     private val nameTempl : String = "_vec%02d"
 
+    private var reductionVarArrayAccesses : Option[IR_ArrayAccess] = None
+
     // init
     pushScope()
 
@@ -241,6 +244,14 @@ private object VectorizeInnermost extends PartialFunction[Node, Transformation.O
     def getAlignedResidue() : Long = {
       alignedResidue
     }
+
+    def setReductionArrayAccess(arrAcc : IR_ArrayAccess) = {
+      reductionVarArrayAccesses = Some(arrAcc)
+    }
+
+    def getReductionArrayAccess() = {
+      reductionVarArrayAccesses
+    }
   }
 
   private def containsVarAcc(node : IR_Node, varName : String) : Boolean = {
@@ -262,8 +273,12 @@ private object VectorizeInnermost extends PartialFunction[Node, Transformation.O
     val ctx = new LoopCtx(itVar, incr)
     var postLoopStmt : IR_Statement = null
     if (reduction.isDefined) {
-      val target = reduction.get.target
+      val target = Duplicate(reduction.get.target)
       val operator = reduction.get.op
+      target match {
+        case arrAcc : IR_ArrayAccess => ctx.setReductionArrayAccess(arrAcc)
+        case _ =>
+      }
 
       val (vecTmp : String, true) = ctx.getName(target)
       val identityElem : IR_Expression =
@@ -602,6 +617,11 @@ private object VectorizeInnermost extends PartialFunction[Node, Transformation.O
 
   private def vectorizeExpr(expr : IR_Expression, ctx : LoopCtx) : IR_Expression = {
     expr match {
+      case arrAcc : IR_ArrayAccess if ctx.getReductionArrayAccess().contains(arrAcc) =>
+        // vec was already added to ctx and declared
+        val (vecTmp : String, false) = ctx.getName(expr)
+        IR_VariableAccess(vecTmp, SIMD_RealDatatype)
+
       // TODO: do not vectorize if base is not aligned?
       case IR_ArrayAccess(base, index, alignedBase) =>
         val (vecTmp : String, njuTmp : Boolean) = ctx.getName(expr)

Compiler/src/exastencils/parallelization/api/cuda/CUDA_AddGlobals.scala

@@ -31,19 +31,16 @@ object CUDA_AddGlobals extends NoTraversalStrategy("Extend globals for CUDA") {
   override def doWork() : Unit = {
     val globals = IR_GlobalCollection.get
 
-    if (!Knowledge.experimental_eliminateCudaContext)
-      globals.variables += IR_VariableDeclaration("CUcontext", "cudaContext")
-    globals.variables += IR_VariableDeclaration("CUdevice", "cudaDevice")
-
     val initFunc = globals.functions.find(_.name == "initGlobals").get.asInstanceOf[IR_Function]
 
     initFunc.body ++= ListBuffer[IR_Statement](
       IR_VariableDeclaration(IR_IntegerDatatype, "deviceCount", 0),
-      "cuDeviceGetCount(&deviceCount)",
+      "cudaGetDeviceCount(&deviceCount)",
       IR_Assert(IR_Lower(Knowledge.cuda_deviceId, "deviceCount"),
         ListBuffer("\"Invalid device id (\"", Knowledge.cuda_deviceId, "\") must be smaller than the number of devices (\"", "deviceCount", "\")\""),
-        IR_FunctionCall("exit", 1)),
-      s"cuDeviceGet(&cudaDevice, ${ Knowledge.cuda_deviceId })")
+        IR_FunctionCall(IR_ExternalFunctionReference("exit"), 1)),
+      s"cudaSetDevice(${ Knowledge.cuda_deviceId })"
+    )
 
     // print device info (name)
     if (!Knowledge.testing_enabled) {
@@ -53,10 +50,6 @@ object CUDA_AddGlobals extends NoTraversalStrategy("Extend globals for CUDA") {
         IR_RawPrint("\"Using CUDA device \"", Knowledge.cuda_deviceId, "\": \"", "devProp.name", "std::endl"))
     }
 
-    // create context
-    if (!Knowledge.experimental_eliminateCudaContext)
-      initFunc.body += "cuCtxCreate(&cudaContext, 0, cudaDevice)"
-
     // set L1 cache and shared memory configuration for this device
     if (Knowledge.cuda_useSharedMemory)
       initFunc.body += "cudaDeviceSetCacheConfig(cudaFuncCachePreferShared)"

Compiler/src/exastencils/parallelization/api/cuda/CUDA_Base.scala

@@ -19,28 +19,8 @@
 package exastencils.parallelization.api.cuda
 
 import exastencils.base.ir._
-import exastencils.config.Knowledge
 import exastencils.prettyprinting.PpStream
 
-/// CUDA_Init
-
-case object CUDA_Init extends CUDA_DeviceStatement {
-  override def prettyprint(out : PpStream) : Unit = {
-    if (!Knowledge.experimental_eliminateCudaContext)
-      out << "cuInit(0);"
-  }
-}
-
-/// CUDA_Finalize
-
-case object CUDA_Finalize extends CUDA_DeviceStatement {
-  override def prettyprint(out : PpStream) : Unit = {
-    // has to be done after all other de-initialization statements
-    if (!Knowledge.experimental_eliminateCudaContext)
-      out << "cuCtxDestroy(cudaContext);"
-  }
-}
-
 /// CUDA_DeviceSynchronize
 
 case class CUDA_DeviceSynchronize() extends CUDA_HostStatement with IR_Expandable {

Compiler/src/exastencils/parallelization/api/cuda/CUDA_Error.scala

@@ -48,16 +48,16 @@ case class CUDA_CheckError(var exp : IR_Expression) extends CUDA_HostStatement w
     def print = IR_RawPrint("\"CUDA error in file (\"", "__FILE__", "\"), line (\"", "__LINE__", "\"): \"", status,
       "\" -> \"", IR_FunctionCall(IR_ExternalFunctionReference("cudaGetErrorString"), status), "std::endl")
 
+    def printAndExit : ListBuffer[IR_Statement] = ListBuffer(print, IR_FunctionCall(IR_ExternalFunctionReference("exit"), 1))
+
     ListBuffer(
       IR_VariableDeclaration(status, exp),
       IR_IfCondition("cudaSuccess" Neq status,
-        print,
+        printAndExit,
         ListBuffer(
           IR_Assignment(status, IR_FunctionCall(IR_ExternalFunctionReference("cudaGetLastError"))),
           IR_IfCondition("cudaSuccess" Neq status,
-            ListBuffer[IR_Statement](
-              print,
-              IR_FunctionCall(IR_ExternalFunctionReference("exit"), 1))
+            printAndExit
           ))))
   }
 }

Compiler/src/exastencils/parallelization/api/cuda/CUDA_ExtractDeviceCode.scala

@@ -114,7 +114,7 @@ object CUDA_ExtractHostAndDeviceCode extends DefaultStrategy("Transform annotate
       loop.getAnnotation(CUDA_Util.CUDA_LOOP_ANNOTATION).contains(CUDA_Util.CUDA_BAND_START) =>
 
       // remove the annotation first to guarantee single application of this transformation.
-      loop.annotate(CUDA_Util.CUDA_LOOP_ANNOTATION)
+      loop.removeAnnotation(CUDA_Util.CUDA_LOOP_ANNOTATION)
 
       val parallelLoops = (x : IR_ForLoop) => {
         x.hasAnnotation(CUDA_Util.CUDA_LOOP_ANNOTATION) &&
@@ -140,11 +140,15 @@ object CUDA_ExtractHostAndDeviceCode extends DefaultStrategy("Transform annotate
 
       val kernelCount = kernelFunctions.counterMap.getOrElse(fctNameCollector.getCurrentName, -1) + 1
 
-      val reduction = loop.parallelization.reduction
+      val reduction = Duplicate(loop.parallelization.reduction)
+      val redTarget = if (reduction.isDefined)
+        Some(Duplicate(reduction.get.target))
+      else
+        None
 
       // local variable for kernels with reductions
       val localTarget = if (reduction.isDefined)
-        Some(IR_VariableAccess(reduction.get.targetName + "_local_" + kernelCount, CUDA_Util.getReductionDatatype(reduction.get.target)))
+        Some(IR_VariableAccess(reduction.get.targetName + "_local_" + kernelCount, CUDA_Util.getReductionDatatype(redTarget.get)))
       else
         None
 
@@ -152,17 +156,17 @@ object CUDA_ExtractHostAndDeviceCode extends DefaultStrategy("Transform annotate
       CUDA_GatherVariableAccesses.clear()
       CUDA_GatherVariableAccesses.kernelCount = kernelCount
       if (reduction.isDefined)
-        CUDA_GatherVariableAccesses.reductionTarget = Some(reduction.get.target)
+        CUDA_GatherVariableAccesses.reductionTarget = redTarget
       CUDA_GatherVariableAccesses.applyStandalone(IR_Scope(loop))
 
       // declare and init local reduction target
       if (localTarget.isDefined) {
         var decl = IR_VariableDeclaration(localTarget.get)
-        var initLocalTarget = CUDA_Util.getReductionDatatype(reduction.get.target) match {
+        var initLocalTarget = CUDA_Util.getReductionDatatype(redTarget.get) match {
           case _ : IR_ScalarDatatype   =>
-            ListBuffer[IR_Statement](IR_Assignment(localTarget.get, reduction.get.target))
+            ListBuffer[IR_Statement](IR_Assignment(localTarget.get, redTarget.get))
           case mat : IR_MatrixDatatype =>
-            reduction.get.target match {
+            redTarget.get match {
               case vAcc : IR_VariableAccess =>
                 IR_GenerateBasicMatrixOperations.loopSetSubmatrixMatPointer(
                   vAcc, localTarget.get, mat.sizeN, mat.sizeM, mat.sizeN, 0, 0).body
@@ -218,7 +222,7 @@ object CUDA_ExtractHostAndDeviceCode extends DefaultStrategy("Transform annotate
       // replace array accesses with accesses to function arguments
       // reduction var is not replaced, but later in IR_HandleReductions
       if (reduction.isDefined)
-        CUDA_ReplaceNonReductionVarArrayAccesses.reductionTarget = Some(reduction.get.target)
+        CUDA_ReplaceNonReductionVarArrayAccesses.reductionTarget = redTarget
       else
         CUDA_ReplaceNonReductionVarArrayAccesses.reductionTarget = None
       CUDA_ReplaceNonReductionVarArrayAccesses.applyStandalone(IR_Scope(kernelBody))

Compiler/src/exastencils/parallelization/api/cuda/CUDA_GatherFieldAccess.scala

@@ -23,6 +23,7 @@ import scala.collection.mutable._
 import exastencils.base.ir._
 import exastencils.core.collectors.Collector
 import exastencils.datastructures._
+import exastencils.domain.ir.IR_IV_NeighborFragmentIdx
 import exastencils.field.ir._
 import exastencils.logger.Logger
 
@@ -85,6 +86,12 @@ class CUDA_GatherFieldAccess extends Collector {
           }
         }
 
+        // also consider neighbor fragment accesses
+        access.fragIdx match {
+          case neigh : IR_IV_NeighborFragmentIdx => identifier += s"_n${ neigh.neighIdx }"
+          case _                                 =>
+        }
+
         if (isRead)
           fieldAccesses.put("read_" + identifier, access)
         if (isWrite)

Compiler/src/exastencils/parallelization/api/cuda/CUDA_HandleFieldAccessLike.scala

@@ -25,6 +25,7 @@ import exastencils.base.ir._
 import exastencils.config._
 import exastencils.datastructures.Transformation._
 import exastencils.datastructures._
+import exastencils.domain.ir.IR_IV_NeighborFragmentIdx
 import exastencils.field.ir._
 import exastencils.optimization.ir.IR_SimplifyExpression
 
@@ -49,6 +50,12 @@ object CUDA_GatherFieldAccessLike extends QuietDefaultStrategy("Gather local Fie
       }
     }
 
+    // also consider neighbor fragment accesses
+    access.fragIdx match {
+      case neigh : IR_IV_NeighborFragmentIdx => identifier += s"_n${ neigh.neighIdx }"
+      case _                                 =>
+    }
+
     identifier
   }
 

Compiler/src/exastencils/parallelization/api/cuda/CUDA_HandleFragmentLoopsWithReduction.scala

@@ -38,17 +38,18 @@ case class CUDA_HandleFragmentLoopsWithReduction(
 
   val iter = IR_LoopOverFragments.defIt
 
-  val redTarget = reduction.target
+  val red = Duplicate(reduction)
+  val redTarget = Duplicate(red.target)
   val reductionDt = CUDA_Util.getReductionDatatype(redTarget)
 
-  val counter = CUDA_HandleFragmentLoopsWithReduction.getReductionCounter(reduction.targetName)
+  val counter = CUDA_HandleFragmentLoopsWithReduction.getReductionCounter(red.targetName)
 
   val copies = {
     val innerDt = reductionDt match {
       case scalar : IR_ScalarDatatype => scalar
       case hodt : IR_HigherDimensionalDatatype => IR_ArrayDatatype(hodt.resolveBaseDatatype, hodt.getSizeArray.product)
     }
-    IR_VariableAccess(reduction.targetName + "_" + counter, IR_ArrayDatatype(innerDt, Knowledge.domain_numFragmentsPerBlock))
+    IR_VariableAccess(red.targetName + "_fragCpy" + counter, IR_ArrayDatatype(innerDt, Knowledge.domain_numFragmentsPerBlock))
   }
   val currCopy = IR_ArrayAccess(copies, iter)
 
@@ -80,11 +81,17 @@ case class CUDA_HandleFragmentLoopsWithReduction(
       matrixAssignment("std::copy", redTarget, currCopy, hodt.getSizeArray.product)
   }
 
-  def initCopies() = ListBuffer(
-    IR_VariableDeclaration(copies), // declare copies
-    IR_LoopOverFragments( // init copies
-      copyReductionTarget()),
-    resetReductionTarget()) // reset initial value as it is already in the copies
+  def initCopies() =  {
+    val declCopies = IR_VariableDeclaration(copies)
+    val initCopies = IR_LoopOverFragments(
+      copyReductionTarget()).expandSpecial().inner
+    val resetRedTarget = resetReductionTarget() // reset initial value as it is already in the copies
+
+    ListBuffer(
+      declCopies,
+      initCopies,
+      resetRedTarget)
+  }
 
   def finalizeReduction(body : ListBuffer[IR_Statement]) = {
     // finalize reduction
@@ -99,10 +106,10 @@ case class CUDA_HandleFragmentLoopsWithReduction(
         val src = IR_ArrayAccess(currCopy, idx)
         IR_ForLoop(IR_VariableDeclaration(i, IR_IntegerConstant(0)), IR_Lower(i, mat.sizeM), IR_PreIncrement(i), ListBuffer[IR_Statement](
           IR_ForLoop(IR_VariableDeclaration(j, 0), IR_Lower(j, mat.sizeN), IR_PreIncrement(j), ListBuffer[IR_Statement](
-            IR_Assignment(dst, IR_BinaryOperators.createExpression(reduction.op, dst, src))))))
+            IR_Assignment(dst, IR_BinaryOperators.createExpression(red.op, dst, src))))))
 
       case _ : IR_ScalarDatatype =>
-        IR_Assignment(redTarget, IR_BinaryOperators.createExpression(reduction.op, redTarget, currCopy))
+        IR_Assignment(redTarget, IR_BinaryOperators.createExpression(red.op, redTarget, currCopy))
     }
 
     body :+ assign
@@ -110,21 +117,21 @@ case class CUDA_HandleFragmentLoopsWithReduction(
 
   def replaceAccesses(body : ListBuffer[IR_Statement]) = {
     // replace occurrences
-    CUDA_ReplaceReductionAccesses.redTarget = redTarget
-    CUDA_ReplaceReductionAccesses.replacement = currCopy
+    CUDA_ReplaceReductionAccesses.redTarget = Duplicate(redTarget)
+    CUDA_ReplaceReductionAccesses.replacement = Duplicate(currCopy)
     CUDA_ReplaceReductionAccesses.applyStandalone(IR_Scope(body))
   }
 
   def addHandling(loop : IR_ForLoop) = {
     replaceAccesses(loop.body)
     loop.body = finalizeReduction(loop.body)
-    initCopies() :+ loop
+    initCopies() :+ Duplicate(loop)
   }
 
   def addHandling(loop : IR_LoopOverFragments) = {
     replaceAccesses(loop.body)
     loop.body = finalizeReduction(loop.body)
-    initCopies() :+ loop
+    initCopies() :+ Duplicate(loop)
   }
 
   override def expand() : OutputType = {

Compiler/src/exastencils/parallelization/api/cuda/CUDA_Kernel.scala

@@ -593,12 +593,12 @@ case class CUDA_Kernel(
     var body = ListBuffer[IR_Statement]()
 
     if (reduction.isDefined) {
-      def target = reduction.get.target
-      def resultDt = CUDA_Util.getReductionDatatype(target)
-      def baseDt = resultDt.resolveBaseDatatype
+      val target = Duplicate(reduction.get.target)
+      val resultDt = CUDA_Util.getReductionDatatype(target)
+      val baseDt = resultDt.resolveBaseDatatype
 
-      def bufSize = requiredThreadsPerDim.product
-      def bufAccess = CUDA_ReductionDeviceData(bufSize, resultDt)
+      val bufSize = requiredThreadsPerDim.product
+      val bufAccess = CUDA_ReductionDeviceData(bufSize, resultDt)
       var callArgsReduction = ListBuffer[IR_Expression](bufAccess, bufSize)
 
       body += CUDA_Memset(bufAccess, 0, bufSize, resultDt)
@@ -621,7 +621,7 @@ case class CUDA_Kernel(
           IR_Return(Some(callDefaultReductionKernel))
       })
 
-      CUDA_KernelFunctions.get.requiredRedKernels += Tuple2(reduction.get.op, target) // request reduction kernel and wrapper
+      CUDA_KernelFunctions.get.requiredRedKernels += Tuple2(reduction.get.op, Duplicate(target)) // request reduction kernel and wrapper
     } else {
       body += CUDA_FunctionCall(getKernelFctName, callArgs, numBlocksPerDim, numThreadsPerBlock)
     }

Compiler/src/exastencils/parallelization/api/cuda/CUDA_Memory.scala

@@ -25,6 +25,8 @@ import exastencils.base.ir._
 import exastencils.baseExt.ir._
 import exastencils.communication.ir._
 import exastencils.config.Knowledge
+import exastencils.datastructures.DefaultStrategy
+import exastencils.datastructures.Transformation
 import exastencils.datastructures.Transformation.Output
 import exastencils.field.ir._
 
@@ -33,7 +35,37 @@ import exastencils.field.ir._
 case class CUDA_Allocate(var pointer : IR_Expression, var numElements : IR_Expression, var datatype : IR_Datatype) extends CUDA_HostStatement with IR_Expandable {
   override def expand() : Output[IR_Statement] = {
     CUDA_CheckError(
-      IR_FunctionCall("cudaMalloc",
+      IR_FunctionCall(IR_ExternalFunctionReference("cudaMalloc"),
+        IR_Cast(IR_PointerDatatype(IR_PointerDatatype(IR_UnitDatatype)), IR_AddressOf(pointer)),
+        numElements * IR_SizeOf(datatype)))
+  }
+}
+
+/// CUDA_AllocateHost
+
+case class CUDA_AllocateHost(var pointer : IR_Expression, var numElements : IR_Expression, var datatype : IR_Datatype) extends CUDA_HostStatement with IR_Expandable {
+  override def expand() : Output[IR_Statement] = {
+    if (Knowledge.cuda_useZeroCopy) {
+      CUDA_CheckError(
+        IR_FunctionCall(IR_ExternalFunctionReference("cudaHostAlloc"),
+          IR_Cast(IR_PointerDatatype(IR_PointerDatatype(IR_UnitDatatype)), IR_AddressOf(pointer)),
+          numElements * IR_SizeOf(datatype),
+          "cudaHostAllocMapped"))
+    } else {
+      CUDA_CheckError(
+        IR_FunctionCall(IR_ExternalFunctionReference("cudaMallocHost"),
+          IR_Cast(IR_PointerDatatype(IR_PointerDatatype(IR_UnitDatatype)), IR_AddressOf(pointer)),
+          numElements * IR_SizeOf(datatype)))
+    }
+  }
+}
+
+/// CUDA_AllocateManaged
+
+case class CUDA_AllocateManaged(var pointer : IR_Expression, var numElements : IR_Expression, var datatype : IR_Datatype) extends CUDA_HostStatement with IR_Expandable {
+  override def expand() : Output[IR_Statement] = {
+    CUDA_CheckError(
+      IR_FunctionCall(IR_ExternalFunctionReference("cudaMallocManaged"),
         IR_Cast(IR_PointerDatatype(IR_PointerDatatype(IR_UnitDatatype)), IR_AddressOf(pointer)),
         numElements * IR_SizeOf(datatype)))
   }
@@ -42,19 +74,38 @@ case class CUDA_Allocate(var pointer : IR_Expression, var numElements : IR_Expre
 /// CUDA_Free
 
 case class CUDA_Free(var pointer : IR_Expression) extends CUDA_HostStatement with IR_Expandable {
-  override def expand() = IR_ExpressionStatement(IR_FunctionCall("cudaFree", pointer))
+  override def expand() = IR_ExpressionStatement(IR_FunctionCall(IR_ExternalFunctionReference("cudaFree"), pointer))
+}
+
+/// CUDA_FreeHost
+
+case class CUDA_FreeHost(var pointer : IR_Expression) extends CUDA_HostStatement with IR_Expandable {
+  override def expand() = IR_ExpressionStatement(IR_FunctionCall(IR_ExternalFunctionReference("cudaFreeHost"), pointer))
 }
 
 /// CUDA_Memcpy
 
 case class CUDA_Memcpy(var dest : IR_Expression, var src : IR_Expression, var sizeInBytes : IR_Expression, var direction : String) extends CUDA_HostStatement with IR_Expandable {
-  override def expand() = CUDA_CheckError(IR_FunctionCall("cudaMemcpy", dest, src, sizeInBytes, direction))
+  override def expand() = CUDA_CheckError(IR_FunctionCall(IR_ExternalFunctionReference("cudaMemcpy"), dest, src, sizeInBytes, direction))
+}
+
+/// CUDA_MemPrefetch
+
+case class CUDA_MemPrefetch(var pointer : IR_Expression, var sizeInBytes : IR_Expression, var target : String) extends CUDA_HostStatement with IR_Expandable {
+  override def expand() = CUDA_CheckError(IR_FunctionCall(IR_ExternalFunctionReference("cudaMemPrefetchAsync "), pointer, sizeInBytes, target))
 }
 
 /// CUDA_Memset
 
 case class CUDA_Memset(var data : IR_Expression, var value : IR_Expression, var numElements : IR_Expression, var datatype : IR_Datatype) extends CUDA_HostStatement with IR_Expandable {