Merge branch 'devel/adapt_alloc' into 'master'

Add option for more compact buffer allocation

See merge request exastencils/exastencils!90

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

@@ -144,6 +144,9 @@ object IR_DefaultLayerHandler extends IR_LayerHandler {
     else
       IR_Expand.doUntilDone()
 
+    if (Knowledge.experimental_compactBufferAllocation)
+      IR_AdaptAllocateDataFunction.apply()
+
     // HACK: create discr_h* again if there are no multigrid level and the field size was defined explicitly
     //   currently this works only if all fields are equally sized
     if (Knowledge.domain_rect_generate && Knowledge.maxLevel <= 0) {

Compiler/src/exastencils/config/Knowledge.scala

@@ -672,6 +672,10 @@ object Knowledge {
 
   /// experimental features
 
+  // compact buffer allocation
+  var experimental_compactBufferAllocation : Boolean = false
+  var experimental_compactBufferAllocationSize : Int = 0 // 0: automatic, else: user-defined
+
   // generates call stacks for all employed timers
   var experimental_disableIterationOffsets : Boolean = false
   var experimental_bc_checkOnlyMainAxis : Boolean = true

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

@@ -21,7 +21,11 @@ package exastencils.globals.ir
 import scala.collection.mutable.ListBuffer
 
 import exastencils.base.ir._
+import exastencils.base.ir.IR_ImplicitConversion._
 import exastencils.communication.NeighborInfo
+import exastencils.config.Knowledge
+import exastencils.datastructures.DefaultStrategy
+import exastencils.datastructures.Transformation
 import exastencils.field.ir._
 
 /// IR_AllocateDataFunction
@@ -30,6 +34,26 @@ object IR_AllocateDataFunction {
   val fctName = "setupBuffers"
 }
 
+object IR_AdaptAllocateDataFunction extends DefaultStrategy("Enable closely spaced allocations") {
+  this += Transformation("Add workaround", {
+    case func : IR_Function if func.name == IR_AllocateDataFunction.fctName =>
+      if (!IR_GlobalCollection.get.externalDependencies.contains("malloc.h"))
+        IR_GlobalCollection.get.externalDependencies += "malloc.h"
+
+      val maxFieldSize = IR_FieldCollection.objects.map(field => (0 until field.layout.numDimsData).map(d =>
+        field.layout.defTotal(d)).product * field.resolveBaseDatatype.typicalByteSize).max
+      val allocSize = if (Knowledge.experimental_compactBufferAllocationSize > 0)
+        Knowledge.experimental_compactBufferAllocationSize
+      else
+        math.ceil(maxFieldSize / 1024.0).toInt * 1024
+
+      func.body.prepend(IR_FunctionCall(IR_ExternalFunctionReference("mallopt"), "M_MMAP_THRESHOLD", allocSize))
+
+      func
+  })
+}
+
+
 // TODO: split to separate functions for (host) fields, communication buffers and device data
 case class IR_AllocateDataFunction(
     var fields : ListBuffer[IR_Field],