merged new hadrons interface

.gitignore

@@ -93,6 +93,7 @@ build*/*
 *.xcodeproj/*
 build.sh
 .vscode
+*.code-workspace
 
 # Eigen source #
 ################
@@ -122,4 +123,3 @@ make-bin-BUCK.sh
 #####################
 lib/qcd/spin/gamma-gen/*.h
 lib/qcd/spin/gamma-gen/*.cc
-

TODO

@@ -1,20 +1,36 @@
 TODO:
 ---------------
 
-Large item work list:
+Code item work list
+
+a) namespaces & indentation
+ GRID_BEGIN_NAMESPACE();
+ GRID_END_NAMESPACE();
+-- delete QCD namespace
+
+b) GPU branch
+- start branch
+- Increase Macro use in core library support; prepare for change
+- Audit volume of "device" code
+- Virtual function audit
+- Start port once Nvidia box is up
+- Cut down volume of code for first port? How?
+
+Physics item work list:
 
 1)- BG/Q port and check ; Andrew says ok.
-2)- Christoph's local basis expansion Lanczos
+2)- Consistent linear solver flop count/rate -- PARTIAL, time but no flop/s yet
---
+3)- Physical propagator interface
-3a)- RNG I/O in ILDG/SciDAC (minor)
+4)- Multigrid Wilson and DWF, compare to other Multigrid implementations
-3b)- Precision conversion and sort out localConvert      <-- partial/easy
+5)- HDCR resume
-3c)- Consistent linear solver flop count/rate -- PARTIAL, time but no flop/s yet
-4)- Physical propagator interface
-5)- Conserved currents
-6)- Multigrid Wilson and DWF, compare to other Multigrid implementations
-7)- HDCR resume
 
+----------------------------
 Recent DONE 
+-- RNG I/O in ILDG/SciDAC (minor) 
+-- Precision conversion and sort out localConvert      <-- partial/easy
+-- Conserved currents (Andrew)
+-- Split grid
+-- Christoph's local basis expansion Lanczos
 -- MultiRHS with spread out extra dim -- Go through filesystem with SciDAC I/O ; <-- DONE ; bmark cori
 -- Lanczos Remove DenseVector, DenseMatrix; Use Eigen instead. <-- DONE
 -- GaugeFix into central location                      <-- DONE

benchmarks/Benchmark_comms.cc

@@ -106,7 +106,7 @@ int main (int argc, char ** argv)
       for(int i=0;i<Nloop;i++){
       double start=usecond();
 
-	std::vector<CartesianCommunicator::CommsRequest_t> requests;
+	std::vector<CommsRequest_t> requests;
 
 	ncomm=0;
 	for(int mu=0;mu<4;mu++){
@@ -202,7 +202,7 @@ int main (int argc, char ** argv)
 	    int recv_from_rank;
 	    
 	    {
-	      std::vector<CartesianCommunicator::CommsRequest_t> requests;
+	      std::vector<CommsRequest_t> requests;
 	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
 	      Grid.SendToRecvFromBegin(requests,
 				       (void *)&xbuf[mu][0],
@@ -215,7 +215,7 @@ int main (int argc, char ** argv)
 
 	    comm_proc = mpi_layout[mu]-1;
 	    {
-	      std::vector<CartesianCommunicator::CommsRequest_t> requests;
+	      std::vector<CommsRequest_t> requests;
 	      Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
 	      Grid.SendToRecvFromBegin(requests,
 				       (void *)&xbuf[mu+4][0],
@@ -290,7 +290,7 @@ int main (int argc, char ** argv)
 	dbytes=0;
 	ncomm=0;
 
-	std::vector<CartesianCommunicator::CommsRequest_t> requests;
+	std::vector<CommsRequest_t> requests;
 
 	for(int mu=0;mu<4;mu++){
 	
@@ -383,7 +383,7 @@ int main (int argc, char ** argv)
       for(int i=0;i<Nloop;i++){
 	double start=usecond();
 
-	std::vector<CartesianCommunicator::CommsRequest_t> requests;
+	std::vector<CommsRequest_t> requests;
 	dbytes=0;
 	ncomm=0;
 	for(int mu=0;mu<4;mu++){
@@ -481,7 +481,7 @@ int main (int argc, char ** argv)
       for(int i=0;i<Nloop;i++){
 	double start=usecond();
 
-	std::vector<CartesianCommunicator::CommsRequest_t> requests;
+	std::vector<CommsRequest_t> requests;
 	dbytes=0;
 	ncomm=0;
 

bootstrap.sh

@@ -3,9 +3,7 @@
 EIGEN_URL='http://bitbucket.org/eigen/eigen/get/3.3.3.tar.bz2'
 
 echo "-- deploying Eigen source..."
-wget ${EIGEN_URL} --no-check-certificate
+wget ${EIGEN_URL} --no-check-certificate && ./scripts/update_eigen.sh `basename ${EIGEN_URL}` && rm `basename ${EIGEN_URL}`
-./scripts/update_eigen.sh `basename ${EIGEN_URL}`
-rm `basename ${EIGEN_URL}`
 
 echo '-- generating Make.inc files...'
 ./scripts/filelist

configure.ac

@@ -337,15 +337,11 @@ case ${ac_PRECISION} in
 esac
 
 ######################  Shared memory allocation technique under MPI3
-AC_ARG_ENABLE([shm],[AC_HELP_STRING([--enable-shm=shmget|shmopen|hugetlbfs],
+AC_ARG_ENABLE([shm],[AC_HELP_STRING([--enable-shm=shmopen|hugetlbfs],
               [Select SHM allocation technique])],[ac_SHM=${enable_shm}],[ac_SHM=shmopen])
 
 case ${ac_SHM} in
 
-     shmget)
-     AC_DEFINE([GRID_MPI3_SHMGET],[1],[GRID_MPI3_SHMGET] )
-     ;;
-
      shmopen)
      AC_DEFINE([GRID_MPI3_SHMOPEN],[1],[GRID_MPI3_SHMOPEN] )
      ;;
@@ -367,7 +363,7 @@ AC_ARG_ENABLE([shmpath],[AC_HELP_STRING([--enable-shmpath=path],
 AC_DEFINE_UNQUOTED([GRID_SHM_PATH],["$ac_SHMPATH"],[Path to a hugetlbfs filesystem for MMAPing])
 
 ############### communication type selection
-AC_ARG_ENABLE([comms],[AC_HELP_STRING([--enable-comms=none|mpi|mpi-auto|mpi3|mpi3-auto|shmem],
+AC_ARG_ENABLE([comms],[AC_HELP_STRING([--enable-comms=none|mpi|mpi-auto],
               [Select communications])],[ac_COMMS=${enable_comms}],[ac_COMMS=none])
 
 case ${ac_COMMS} in
@@ -375,22 +371,10 @@ case ${ac_COMMS} in
         AC_DEFINE([GRID_COMMS_NONE],[1],[GRID_COMMS_NONE] )
         comms_type='none'
      ;;
-     mpi3*)
+     mpi*)
         AC_DEFINE([GRID_COMMS_MPI3],[1],[GRID_COMMS_MPI3] )
         comms_type='mpi3'
      ;;
-     mpit)
-        AC_DEFINE([GRID_COMMS_MPIT],[1],[GRID_COMMS_MPIT] )
-        comms_type='mpit'
-     ;;
-     mpi*)
-        AC_DEFINE([GRID_COMMS_MPI],[1],[GRID_COMMS_MPI] )
-        comms_type='mpi'
-     ;;
-     shmem)
-        AC_DEFINE([GRID_COMMS_SHMEM],[1],[GRID_COMMS_SHMEM] )
-        comms_type='shmem'
-     ;;
      *)
         AC_MSG_ERROR([${ac_COMMS} unsupported --enable-comms option]);
      ;;
@@ -550,6 +534,7 @@ AC_CONFIG_FILES(tests/forces/Makefile)
 AC_CONFIG_FILES(tests/hadrons/Makefile)
 AC_CONFIG_FILES(tests/hmc/Makefile)
 AC_CONFIG_FILES(tests/solver/Makefile)
+AC_CONFIG_FILES(tests/lanczos/Makefile)
 AC_CONFIG_FILES(tests/smearing/Makefile)
 AC_CONFIG_FILES(tests/qdpxx/Makefile)
 AC_CONFIG_FILES(tests/testu01/Makefile)

extras/Hadrons/Application.cc

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Application.cc
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -43,6 +42,7 @@ using namespace Hadrons;
 // constructors ////////////////////////////////////////////////////////////////
 Application::Application(void)
 {
+    initLogger();
     LOG(Message) << "Modules available:" << std::endl;
     auto list = ModuleFactory::getInstance().getBuilderList();
     for (auto &m: list)
@@ -73,12 +73,6 @@ Application::Application(const std::string parameterFileName)
     parameterFileName_ = parameterFileName;
 }
 
-// environment shortcut ////////////////////////////////////////////////////////
-Environment & Application::env(void) const
-{
-    return Environment::getInstance();
-}
-
 // access //////////////////////////////////////////////////////////////////////
 void Application::setPar(const Application::GlobalPar &par)
 {
@@ -94,14 +88,13 @@ const Application::GlobalPar & Application::getPar(void)
 // execute /////////////////////////////////////////////////////////////////////
 void Application::run(void)
 {
-    if (!parameterFileName_.empty() and (env().getNModule() == 0))
+    if (!parameterFileName_.empty() and (vm().getNModule() == 0))
     {
         parseParameterFile(parameterFileName_);
     }
-    if (!scheduled_)
+    vm().printContent();
-    {
+    env().printContent();
-        schedule();
+    schedule();
-    }
     printSchedule();
     configLoop();
 }
@@ -124,12 +117,20 @@ void Application::parseParameterFile(const std::string parameterFileName)
     LOG(Message) << "Building application from '" << parameterFileName << "'..." << std::endl;
     read(reader, "parameters", par);
     setPar(par);
-    push(reader, "modules");
+    if (!push(reader, "modules"))
-    push(reader, "module");
+    {
+        HADRON_ERROR(Parsing, "Cannot open node 'modules' in parameter file '" 
+                              + parameterFileName + "'");
+    }
+    if (!push(reader, "module"))
+    {
+        HADRON_ERROR(Parsing, "Cannot open node 'modules/module' in parameter file '" 
+                              + parameterFileName + "'");
+    }
     do
     {
         read(reader, "id", id);
-        env().createModule(id.name, id.type, reader);
+        vm().createModule(id.name, id.type, reader);
     } while (reader.nextElement("module"));
     pop(reader);
     pop(reader);
@@ -139,7 +140,7 @@ void Application::saveParameterFile(const std::string parameterFileName)
 {
     XmlWriter          writer(parameterFileName);
     ObjectId           id;
-    const unsigned int nMod = env().getNModule();
+    const unsigned int nMod = vm().getNModule();
     
     LOG(Message) << "Saving application to '" << parameterFileName << "'..." << std::endl;
     write(writer, "parameters", getPar());
@@ -147,10 +148,10 @@ void Application::saveParameterFile(const std::string parameterFileName)
     for (unsigned int i = 0; i < nMod; ++i)
     {
         push(writer, "module");
-        id.name = env().getModuleName(i);
+        id.name = vm().getModuleName(i);
-        id.type = env().getModule(i)->getRegisteredName();
+        id.type = vm().getModule(i)->getRegisteredName();
         write(writer, "id", id);
-        env().getModule(i)->saveParameters(writer, "options");
+        vm().getModule(i)->saveParameters(writer, "options");
         pop(writer);
     }
     pop(writer);
@@ -158,95 +159,13 @@ void Application::saveParameterFile(const std::string parameterFileName)
 }
 
 // schedule computation ////////////////////////////////////////////////////////
-#define MEM_MSG(size)\
-sizeString((size)*locVol_) << " (" << sizeString(size)  << "/site)"
-
-#define DEFINE_MEMPEAK \
-GeneticScheduler<unsigned int>::ObjFunc memPeak = \
-[this](const std::vector<unsigned int> &program)\
-{\
-    unsigned int memPeak;\
-    bool         msg;\
-    \
-    msg = HadronsLogMessage.isActive();\
-    HadronsLogMessage.Active(false);\
-    env().dryRun(true);\
-    memPeak = env().executeProgram(program);\
-    env().dryRun(false);\
-    env().freeAll();\
-    HadronsLogMessage.Active(true);\
-    \
-    return memPeak;\
-}
-
 void Application::schedule(void)
 {
-    DEFINE_MEMPEAK;
+    if (!scheduled_ and !loadedSchedule_)
-    
-    // build module dependency graph
-    LOG(Message) << "Building module graph..." << std::endl;
-    auto graph = env().makeModuleGraph();
-    auto con = graph.getConnectedComponents();
-    
-    // constrained topological sort using a genetic algorithm
-    LOG(Message) << "Scheduling computation..." << std::endl;
-    LOG(Message) << "               #module= " << graph.size() << std::endl;
-    LOG(Message) << "       population size= " << par_.genetic.popSize << std::endl;
-    LOG(Message) << "       max. generation= " << par_.genetic.maxGen << std::endl;
-    LOG(Message) << "  max. cst. generation= " << par_.genetic.maxCstGen << std::endl;
-    LOG(Message) << "         mutation rate= " << par_.genetic.mutationRate << std::endl;
-    
-    unsigned int                               k = 0, gen, prevPeak, nCstPeak = 0;
-    std::random_device                         rd;
-    GeneticScheduler<unsigned int>::Parameters par;
-    
-    par.popSize      = par_.genetic.popSize;
-    par.mutationRate = par_.genetic.mutationRate;
-    par.seed         = rd();
-    memPeak_         = 0;
-    CartesianCommunicator::BroadcastWorld(0, &(par.seed), sizeof(par.seed));
-    for (unsigned int i = 0; i < con.size(); ++i)
     {
-        GeneticScheduler<unsigned int> scheduler(con[i], memPeak, par);
+        program_   = vm().schedule(par_.genetic);
-        
+        scheduled_ = true;
-        gen = 0;
-        do
-        {
-            LOG(Debug) << "Generation " << gen << ":" << std::endl;
-            scheduler.nextGeneration();
-            if (gen != 0)
-            {
-                if (prevPeak == scheduler.getMinValue())
-                {
-                    nCstPeak++;
-                }
-                else
-                {
-                    nCstPeak = 0;
-                }
-            }
-            
-            prevPeak = scheduler.getMinValue();
-            if (gen % 10 == 0)
-            {
-                LOG(Iterative) << "Generation " << gen << ": "
-                               << MEM_MSG(scheduler.getMinValue()) << std::endl;
-            }
-            
-            gen++;
-        } while ((gen < par_.genetic.maxGen)
-                 and (nCstPeak < par_.genetic.maxCstGen));
-        auto &t = scheduler.getMinSchedule();
-        if (scheduler.getMinValue() > memPeak_)
-        {
-            memPeak_ = scheduler.getMinValue();
-        }
-        for (unsigned int j = 0; j < t.size(); ++j)
-        {
-            program_.push_back(t[j]);
-        }
     }
-    scheduled_ = true;
 }
 
 void Application::saveSchedule(const std::string filename)
@@ -256,21 +175,19 @@ void Application::saveSchedule(const std::string filename)
     
     if (!scheduled_)
     {
-        HADRON_ERROR("Computation not scheduled");
+        HADRON_ERROR(Definition, "Computation not scheduled");
     }
     LOG(Message) << "Saving current schedule to '" << filename << "'..."
                  << std::endl;
     for (auto address: program_)
     {
-        program.push_back(env().getModuleName(address));
+        program.push_back(vm().getModuleName(address));
     }
     write(writer, "schedule", program);
 }
 
 void Application::loadSchedule(const std::string filename)
 {
-    DEFINE_MEMPEAK;
-    
     TextReader               reader(filename);
     std::vector<std::string> program;
     
@@ -280,24 +197,24 @@ void Application::loadSchedule(const std::string filename)
     program_.clear();
     for (auto &name: program)
     {
-        program_.push_back(env().getModuleAddress(name));
+        program_.push_back(vm().getModuleAddress(name));
     }
-    scheduled_ = true;
+    loadedSchedule_ = true;
-    memPeak_   = memPeak(program_);
 }
 
 void Application::printSchedule(void)
 {
     if (!scheduled_)
     {
-        HADRON_ERROR("Computation not scheduled");
+        HADRON_ERROR(Definition, "Computation not scheduled");
     }
-    LOG(Message) << "Schedule (memory peak: " << MEM_MSG(memPeak_) << "):"
+    auto peak = vm().memoryNeeded(program_);
+    LOG(Message) << "Schedule (memory needed: " << sizeString(peak) << "):"
                  << std::endl;
     for (unsigned int i = 0; i < program_.size(); ++i)
     {
         LOG(Message) << std::setw(4) << i + 1 << ": "
-                     << env().getModuleName(program_[i]) << std::endl;
+                     << vm().getModuleName(program_[i]) << std::endl;
     }
 }
 
@@ -310,8 +227,8 @@ void Application::configLoop(void)
     {
         LOG(Message) << BIG_SEP << " Starting measurement for trajectory " << t
                      << " " << BIG_SEP << std::endl;
-        env().setTrajectory(t);
+        vm().setTrajectory(t);
-        env().executeProgram(program_);
+        vm().executeProgram(program_);
     }
     LOG(Message) << BIG_SEP << " End of measurement " << BIG_SEP << std::endl;
     env().freeAll();

extras/Hadrons/Application.hpp

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Application.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -31,8 +30,7 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #define Hadrons_Application_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
-#include <Grid/Hadrons/Environment.hpp>
+#include <Grid/Hadrons/VirtualMachine.hpp>
-#include <Grid/Hadrons/ModuleFactory.hpp>
 #include <Grid/Hadrons/Modules.hpp>
 
 BEGIN_HADRONS_NAMESPACE
@@ -51,25 +49,13 @@ public:
                                         unsigned int, end,
                                         unsigned int, step);
     };
-    class GeneticPar: Serializable
-    {
-    public:
-        GeneticPar(void):
-            popSize{20}, maxGen{1000}, maxCstGen{100}, mutationRate{.1} {};
-    public:
-        GRID_SERIALIZABLE_CLASS_MEMBERS(GeneticPar,
-                                        unsigned int, popSize,
-                                        unsigned int, maxGen,
-                                        unsigned int, maxCstGen,
-                                        double      , mutationRate);
-    };
     class GlobalPar: Serializable
     {
     public:
         GRID_SERIALIZABLE_CLASS_MEMBERS(GlobalPar,
-                                        TrajRange,   trajCounter,
+                                        TrajRange,                  trajCounter,
-                                        GeneticPar,  genetic,
+                                        VirtualMachine::GeneticPar, genetic,
-                                        std::string, seed);
+                                        std::string,                seed);
     };
 public:
     // constructors
@@ -100,14 +86,15 @@ public:
     void configLoop(void);
 private:
     // environment shortcut
-    Environment & env(void) const;
+    DEFINE_ENV_ALIAS;
+    // virtual machine shortcut
+    DEFINE_VM_ALIAS;
 private:
-    long unsigned int         locVol_;
+    long unsigned int       locVol_;
-    std::string               parameterFileName_{""};
+    std::string             parameterFileName_{""};
-    GlobalPar                 par_;
+    GlobalPar               par_;
-    std::vector<unsigned int> program_;
+    VirtualMachine::Program program_;
-    Environment::Size         memPeak_;
+    bool                    scheduled_{false}, loadedSchedule_{false};
-    bool                      scheduled_{false};
 };
 
 /******************************************************************************
@@ -117,14 +104,16 @@ private:
 template <typename M>
 void Application::createModule(const std::string name)
 {
-    env().createModule<M>(name);
+    vm().createModule<M>(name);
+    scheduled_ = false;
 }
 
 template <typename M>
 void Application::createModule(const std::string name,
                                const typename M::Par &par)
 {
-    env().createModule<M>(name, par);
+    vm().createModule<M>(name, par);
+    scheduled_ = false;
 }
 
 END_HADRONS_NAMESPACE

extras/Hadrons/Environment.cc

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Environment.cc
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -35,6 +34,9 @@ using namespace Grid;
 using namespace QCD;
 using namespace Hadrons;
 
+#define ERROR_NO_ADDRESS(address)\
+HADRON_ERROR(Definition, "no object with address " + std::to_string(address));
+
 /******************************************************************************
  *                       Environment implementation                           *
  ******************************************************************************/
@@ -56,28 +58,6 @@ Environment::Environment(void)
     rng4d_.reset(new GridParallelRNG(grid4d_.get()));
 }
 
-// dry run /////////////////////////////////////////////////////////////////////
-void Environment::dryRun(const bool isDry)
-{
-    dryRun_ = isDry;
-}
-
-bool Environment::isDryRun(void) const
-{
-    return dryRun_;
-}
-
-// trajectory number ///////////////////////////////////////////////////////////
-void Environment::setTrajectory(const unsigned int traj)
-{
-    traj_ = traj;
-}
-
-unsigned int Environment::getTrajectory(void) const
-{
-    return traj_;
-}
-
 // grids ///////////////////////////////////////////////////////////////////////
 void Environment::createGrid(const unsigned int Ls)
 {
@@ -105,7 +85,7 @@ GridCartesian * Environment::getGrid(const unsigned int Ls) const
     }
     catch(std::out_of_range &)
     {
-        HADRON_ERROR("no grid with Ls= " << Ls);
+        HADRON_ERROR(Definition, "no grid with Ls= " + std::to_string(Ls));
     }
 }
 
@@ -124,7 +104,7 @@ GridRedBlackCartesian * Environment::getRbGrid(const unsigned int Ls) const
     }
     catch(std::out_of_range &)
     {
-        HADRON_ERROR("no red-black 5D grid with Ls= " << Ls);
+        HADRON_ERROR(Definition, "no red-black 5D grid with Ls= " + std::to_string(Ls));
     }
 }
 
@@ -143,6 +123,11 @@ int Environment::getDim(const unsigned int mu) const
     return dim_[mu];
 }
 
+unsigned long int Environment::getLocalVolume(void) const
+{
+    return locVol_;
+}
+
 // random number generator /////////////////////////////////////////////////////
 void Environment::setSeed(const std::vector<int> &seed)
 {
@@ -154,291 +139,6 @@ GridParallelRNG * Environment::get4dRng(void) const
     return rng4d_.get();
 }
 
-// module management ///////////////////////////////////////////////////////////
-void Environment::pushModule(Environment::ModPt &pt)
-{
-    std::string name = pt->getName();
-    
-    if (!hasModule(name))
-    {
-        std::vector<unsigned int> inputAddress;
-        unsigned int              address;
-        ModuleInfo                m;
-        
-        m.data = std::move(pt);
-        m.type = typeIdPt(*m.data.get());
-        m.name = name;
-        auto input  = m.data->getInput();
-        for (auto &in: input)
-        {
-            if (!hasObject(in))
-            {
-                addObject(in , -1);
-            }
-            m.input.push_back(objectAddress_[in]);
-        }
-        auto output = m.data->getOutput();
-        module_.push_back(std::move(m));
-        address              = static_cast<unsigned int>(module_.size() - 1);
-        moduleAddress_[name] = address;
-        for (auto &out: output)
-        {
-            if (!hasObject(out))
-            {
-                addObject(out, address);
-            }
-            else
-            {
-                if (object_[objectAddress_[out]].module < 0)
-                {
-                    object_[objectAddress_[out]].module = address;
-                }
-                else
-                {
-                    HADRON_ERROR("object '" + out
-                                 + "' is already produced by module '"
-                                 + module_[object_[getObjectAddress(out)].module].name
-                                 + "' (while pushing module '" + name + "')");
-                }
-            }
-        }
-    }
-    else
-    {
-        HADRON_ERROR("module '" + name + "' already exists");
-    }
-}
-
-unsigned int Environment::getNModule(void) const
-{
-    return module_.size();
-}
-
-void Environment::createModule(const std::string name, const std::string type,
-                               XmlReader &reader)
-{
-    auto &factory = ModuleFactory::getInstance();
-    auto pt       = factory.create(type, name);
-    
-    pt->parseParameters(reader, "options");
-    pushModule(pt);
-}
-
-ModuleBase * Environment::getModule(const unsigned int address) const
-{
-    if (hasModule(address))
-    {
-        return module_[address].data.get();
-    }
-    else
-    {
-        HADRON_ERROR("no module with address " + std::to_string(address));
-    }
-}
-
-ModuleBase * Environment::getModule(const std::string name) const
-{
-    return getModule(getModuleAddress(name));
-}
-
-unsigned int Environment::getModuleAddress(const std::string name) const
-{
-    if (hasModule(name))
-    {
-        return moduleAddress_.at(name);
-    }
-    else
-    {
-        HADRON_ERROR("no module with name '" + name + "'");
-    }
-}
-
-std::string Environment::getModuleName(const unsigned int address) const
-{
-    if (hasModule(address))
-    {
-        return module_[address].name;
-    }
-    else
-    {
-        HADRON_ERROR("no module with address " + std::to_string(address));
-    }
-}
-
-std::string Environment::getModuleType(const unsigned int address) const
-{
-    if (hasModule(address))
-    {
-        return typeName(module_[address].type);
-    }
-    else
-    {
-        HADRON_ERROR("no module with address " + std::to_string(address));
-    }
-}
-
-std::string Environment::getModuleType(const std::string name) const
-{
-    return getModuleType(getModuleAddress(name));
-}
-
-std::string Environment::getModuleNamespace(const unsigned int address) const
-{
-    std::string type = getModuleType(address), ns;
-    
-    auto pos2 = type.rfind("::");
-    auto pos1 = type.rfind("::", pos2 - 2);
-    
-    return type.substr(pos1 + 2, pos2 - pos1 - 2);
-}
-
-std::string Environment::getModuleNamespace(const std::string name) const
-{
-    return getModuleNamespace(getModuleAddress(name));
-}
-
-bool Environment::hasModule(const unsigned int address) const
-{
-    return (address < module_.size());
-}
-
-bool Environment::hasModule(const std::string name) const
-{
-    return (moduleAddress_.find(name) != moduleAddress_.end());
-}
-
-Graph<unsigned int> Environment::makeModuleGraph(void) const
-{
-    Graph<unsigned int> moduleGraph;
-    
-    for (unsigned int i = 0; i < module_.size(); ++i)
-    {
-        moduleGraph.addVertex(i);
-        for (auto &j: module_[i].input)
-        {
-            moduleGraph.addEdge(object_[j].module, i);
-        }
-    }
-    
-    return moduleGraph;
-}
-
-#define BIG_SEP "==============="
-#define SEP     "---------------"
-#define MEM_MSG(size)\
-sizeString((size)*locVol_) << " (" << sizeString(size)  << "/site)"
-
-Environment::Size
-Environment::executeProgram(const std::vector<unsigned int> &p)
-{
-    Size                                memPeak = 0, sizeBefore, sizeAfter;
-    std::vector<std::set<unsigned int>> freeProg;
-    bool                                continueCollect, nothingFreed;
-    
-    // build garbage collection schedule
-    freeProg.resize(p.size());
-    for (unsigned int i = 0; i < object_.size(); ++i)
-    {
-        auto pred = [i, this](const unsigned int j)
-        {
-            auto &in = module_[j].input;
-            auto it  = std::find(in.begin(), in.end(), i);
-            
-            return (it != in.end()) or (j == object_[i].module);
-        };
-        auto it = std::find_if(p.rbegin(), p.rend(), pred);
-        if (it != p.rend())
-        {
-            freeProg[p.rend() - it - 1].insert(i);
-        }
-    }
-    
-    // program execution
-    for (unsigned int i = 0; i < p.size(); ++i)
-    {
-        // execute module
-        if (!isDryRun())
-        {
-            LOG(Message) << SEP << " Measurement step " << i+1 << "/"
-                         << p.size() << " (module '" << module_[p[i]].name
-                         << "') " << SEP << std::endl;
-        }
-        (*module_[p[i]].data)();
-        sizeBefore = getTotalSize();
-        // print used memory after execution
-        if (!isDryRun())
-        {
-            LOG(Message) << "Allocated objects: " << MEM_MSG(sizeBefore)
-                         << std::endl;
-        }
-        if (sizeBefore > memPeak)
-        {
-            memPeak = sizeBefore;
-        }
-        // garbage collection for step i
-        if (!isDryRun())
-        {
-            LOG(Message) << "Garbage collection..." << std::endl;
-        }
-        nothingFreed = true;
-        do
-        {
-            continueCollect = false;
-            auto toFree = freeProg[i];
-            for (auto &j: toFree)
-            {
-                // continue garbage collection while there are still
-                // objects without owners
-                continueCollect = continueCollect or !hasOwners(j);
-                if(freeObject(j))
-                {
-                    // if an object has been freed, remove it from
-                    // the garbage collection schedule
-                    freeProg[i].erase(j);
-                    nothingFreed = false;
-                }
-            }
-        } while (continueCollect);
-        // any remaining objects in step i garbage collection schedule
-        // is scheduled for step i + 1
-        if (i + 1 < p.size())
-        {
-            for (auto &j: freeProg[i])
-            {
-                freeProg[i + 1].insert(j);
-            }
-        }
-        // print used memory after garbage collection if necessary
-        if (!isDryRun())
-        {
-            sizeAfter = getTotalSize();
-            if (sizeBefore != sizeAfter)
-            {
-                LOG(Message) << "Allocated objects: " << MEM_MSG(sizeAfter)
-                             << std::endl;
-            }
-            else
-            {
-                LOG(Message) << "Nothing to free" << std::endl;
-            }
-        }
-    }
-    
-    return memPeak;
-}
-
-Environment::Size Environment::executeProgram(const std::vector<std::string> &p)
-{
-    std::vector<unsigned int> pAddress;
-    
-    for (auto &n: p)
-    {
-        pAddress.push_back(getModuleAddress(n));
-    }
-    
-    return executeProgram(pAddress);
-}
-
 // general memory management ///////////////////////////////////////////////////
 void Environment::addObject(const std::string name, const int moduleAddress)
 {
@@ -448,46 +148,25 @@ void Environment::addObject(const std::string name, const int moduleAddress)
         
         info.name   = name;
         info.module = moduleAddress;
+        info.data   = nullptr;
         object_.push_back(std::move(info));
         objectAddress_[name] = static_cast<unsigned int>(object_.size() - 1);
     }
     else
     {
-        HADRON_ERROR("object '" + name + "' already exists");
+        HADRON_ERROR(Definition, "object '" + name + "' already exists");
     }
 }
 
-void Environment::registerObject(const unsigned int address,
+void Environment::setObjectModule(const unsigned int objAddress,
-                                 const unsigned int size, const unsigned int Ls)
+                                  const int modAddress)
 {
-    if (!hasRegisteredObject(address))
+    object_[objAddress].module = modAddress;
-    {
-        if (hasObject(address))
-        {
-            object_[address].size         = size;
-            object_[address].Ls           = Ls;
-            object_[address].isRegistered = true;
-        }
-        else
-        {
-            HADRON_ERROR("no object with address " + std::to_string(address));
-        }
-    }
-    else
-    {
-        HADRON_ERROR("object with address " + std::to_string(address)
-                     + " already registered");
-    }
 }
 
-void Environment::registerObject(const std::string name,
+unsigned int Environment::getMaxAddress(void) const
-                                 const unsigned int size, const unsigned int Ls)
 {
-    if (!hasObject(name))
+    return object_.size();
-    {
-        addObject(name);
-    }
-    registerObject(getObjectAddress(name), size, Ls);
 }
 
 unsigned int Environment::getObjectAddress(const std::string name) const
@@ -498,7 +177,7 @@ unsigned int Environment::getObjectAddress(const std::string name) const
     }
     else
     {
-        HADRON_ERROR("no object with name '" + name + "'");
+        HADRON_ERROR(Definition, "no object with name '" + name + "'");
     }
 }
 
@@ -510,13 +189,13 @@ std::string Environment::getObjectName(const unsigned int address) const
     }
     else
     {
-        HADRON_ERROR("no object with address " + std::to_string(address));
+        ERROR_NO_ADDRESS(address);
     }
 }
 
 std::string Environment::getObjectType(const unsigned int address) const
 {
-    if (hasRegisteredObject(address))
+    if (hasObject(address))
     {
         if (object_[address].type)
         {
@@ -527,14 +206,9 @@ std::string Environment::getObjectType(const unsigned int address) const
             return "<no type>";
         }
     }
-    else if (hasObject(address))
-    {
-        HADRON_ERROR("object with address " + std::to_string(address)
-                     + " exists but is not registered");
-    }
     else
     {
-        HADRON_ERROR("no object with address " + std::to_string(address));
+        ERROR_NO_ADDRESS(address);
     }
 }
 
@@ -545,18 +219,13 @@ std::string Environment::getObjectType(const std::string name) const
 
 Environment::Size Environment::getObjectSize(const unsigned int address) const
 {
-    if (hasRegisteredObject(address))
+    if (hasObject(address))
     {
         return object_[address].size;
     }
-    else if (hasObject(address))
-    {
-        HADRON_ERROR("object with address " + std::to_string(address)
-                     + " exists but is not registered");
-    }
     else
     {
-        HADRON_ERROR("no object with address " + std::to_string(address));
+        ERROR_NO_ADDRESS(address);
     }
 }
 
@@ -565,7 +234,24 @@ Environment::Size Environment::getObjectSize(const std::string name) const
     return getObjectSize(getObjectAddress(name));
 }
 
-unsigned int Environment::getObjectModule(const unsigned int address) const
+Environment::Storage Environment::getObjectStorage(const unsigned int address) const
+{
+    if (hasObject(address))
+    {
+        return object_[address].storage;
+    }
+    else
+    {
+        ERROR_NO_ADDRESS(address);
+    }
+}
+
+Environment::Storage Environment::getObjectStorage(const std::string name) const
+{
+    return getObjectStorage(getObjectAddress(name));
+}
+
+int Environment::getObjectModule(const unsigned int address) const
 {
     if (hasObject(address))
     {
@@ -573,29 +259,24 @@ unsigned int Environment::getObjectModule(const unsigned int address) const
     }
     else
     {
-        HADRON_ERROR("no object with address " + std::to_string(address));
+        ERROR_NO_ADDRESS(address);
     }
 }
 
-unsigned int Environment::getObjectModule(const std::string name) const
+int Environment::getObjectModule(const std::string name) const
 {
     return getObjectModule(getObjectAddress(name));
 }
 
 unsigned int Environment::getObjectLs(const unsigned int address) const
 {
-    if (hasRegisteredObject(address))
+    if (hasObject(address))
     {
         return object_[address].Ls;
     }
-    else if (hasObject(address))
-    {
-        HADRON_ERROR("object with address " + std::to_string(address)
-                     + " exists but is not registered");
-    }
     else
     {
-        HADRON_ERROR("no object with address " + std::to_string(address));
+        ERROR_NO_ADDRESS(address);
     }
 }
 
@@ -616,30 +297,6 @@ bool Environment::hasObject(const std::string name) const
     return ((it != objectAddress_.end()) and hasObject(it->second));
 }
 
-bool Environment::hasRegisteredObject(const unsigned int address) const
-{
-    if (hasObject(address))
-    {
-        return object_[address].isRegistered;
-    }
-    else
-    {
-        return false;
-    }
-}
-
-bool Environment::hasRegisteredObject(const std::string name) const
-{
-    if (hasObject(name))
-    {
-        return hasRegisteredObject(getObjectAddress(name));
-    }
-    else
-    {
-        return false;
-    }
-}
-
 bool Environment::hasCreatedObject(const unsigned int address) const
 {
     if (hasObject(address))
@@ -680,92 +337,27 @@ Environment::Size Environment::getTotalSize(void) const
     
     for (auto &o: object_)
     {
-        if (o.isRegistered)
+        size += o.size;
-        {
-            size += o.size;
-        }
     }
     
     return size;
 }
 
-void Environment::addOwnership(const unsigned int owner,
+void Environment::freeObject(const unsigned int address)
-                               const unsigned int property)
 {
-    if (hasObject(property))
+    if (hasCreatedObject(address))
     {
-        object_[property].owners.insert(owner);
+        LOG(Message) << "Destroying object '" << object_[address].name
-    }
+                     << "'" << std::endl;
-    else
-    {
-        HADRON_ERROR("no object with address " + std::to_string(property));
-    }
-    if (hasObject(owner))
-    {
-        object_[owner].properties.insert(property);
-    }
-    else
-    {
-        HADRON_ERROR("no object with address " + std::to_string(owner));
     }
+    object_[address].size = 0;
+    object_[address].type = nullptr;
+    object_[address].data.reset(nullptr);
 }
 
-void Environment::addOwnership(const std::string owner,
+void Environment::freeObject(const std::string name)
-                               const std::string property)
 {
-    addOwnership(getObjectAddress(owner), getObjectAddress(property));
+    freeObject(getObjectAddress(name));
-}
-
-bool Environment::hasOwners(const unsigned int address) const
-{
-    
-    if (hasObject(address))
-    {
-        return (!object_[address].owners.empty());
-    }
-    else
-    {
-        HADRON_ERROR("no object with address " + std::to_string(address));
-    }
-}
-
-bool Environment::hasOwners(const std::string name) const
-{
-    return hasOwners(getObjectAddress(name));
-}
-
-bool Environment::freeObject(const unsigned int address)
-{
-    if (!hasOwners(address))
-    {
-        if (!isDryRun() and object_[address].isRegistered)
-        {
-            LOG(Message) << "Destroying object '" << object_[address].name
-                         << "'" << std::endl;
-        }
-        for (auto &p: object_[address].properties)
-        {
-            object_[p].owners.erase(address);
-        }
-        object_[address].size         = 0;
-        object_[address].Ls           = 0;
-        object_[address].isRegistered = false;
-        object_[address].type         = nullptr;
-        object_[address].owners.clear();
-        object_[address].properties.clear();
-        object_[address].data.reset(nullptr);
-        
-        return true;
-    }
-    else
-    {
-        return false;
-    }
-}
-
-bool Environment::freeObject(const std::string name)
-{
-    return freeObject(getObjectAddress(name));
 }
 
 void Environment::freeAll(void)
@@ -776,18 +368,24 @@ void Environment::freeAll(void)
     }
 }
 
-void Environment::printContent(void)
+void Environment::protectObjects(const bool protect)
 {
-    LOG(Message) << "Modules: " << std::endl;
+    protect_ = protect;
-    for (unsigned int i = 0; i < module_.size(); ++i)
+}
-    {
+
-        LOG(Message) << std::setw(4) << i << ": "
+bool Environment::objectsProtected(void) const
-                     << getModuleName(i) << std::endl;
+{
-    }
+    return protect_;
-    LOG(Message) << "Objects: " << std::endl;
+}
+
+// print environment content ///////////////////////////////////////////////////
+void Environment::printContent(void) const
+{
+    LOG(Debug) << "Objects: " << std::endl;
     for (unsigned int i = 0; i < object_.size(); ++i)
     {
-        LOG(Message) << std::setw(4) << i << ": "
+        LOG(Debug) << std::setw(4) << i << ": "
-                     << getObjectName(i) << std::endl;
+                   << getObjectName(i) << " ("
+                   << sizeString(getObjectSize(i)) << ")" << std::endl;
     }
 }

extras/Hadrons/Environment.hpp

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Environment.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -31,20 +30,12 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #define Hadrons_Environment_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
-#include <Grid/Hadrons/Graph.hpp>
-
-#ifndef SITE_SIZE_TYPE
-#define SITE_SIZE_TYPE unsigned int
-#endif
 
 BEGIN_HADRONS_NAMESPACE
 
 /******************************************************************************
  *                         Global environment                                 *
  ******************************************************************************/
-// forward declaration of Module
-class ModuleBase;
-
 class Object
 {
 public:
@@ -66,123 +57,78 @@ private:
     std::unique_ptr<T> objPt_{nullptr};
 };
 
+#define DEFINE_ENV_ALIAS \
+inline Environment & env(void) const\
+{\
+    return Environment::getInstance();\
+}
+
 class Environment
 {
     SINGLETON(Environment);
 public:
     typedef SITE_SIZE_TYPE                         Size;
-    typedef std::unique_ptr<ModuleBase>            ModPt;
     typedef std::unique_ptr<GridCartesian>         GridPt;
     typedef std::unique_ptr<GridRedBlackCartesian> GridRbPt;
     typedef std::unique_ptr<GridParallelRNG>       RngPt;
-    typedef std::unique_ptr<LatticeBase>           LatticePt;
+    enum class Storage {object, cache, temporary};
 private:
-    struct ModuleInfo
-    {
-        const std::type_info      *type{nullptr};
-        std::string               name;
-        ModPt                     data{nullptr};
-        std::vector<unsigned int> input;
-    };
     struct ObjInfo
     {
         Size                    size{0};
+        Storage                 storage{Storage::object};
         unsigned int            Ls{0};
-        bool                    isRegistered{false};
         const std::type_info    *type{nullptr};
         std::string             name;
         int                     module{-1};
-        std::set<unsigned int>  owners, properties;
         std::unique_ptr<Object> data{nullptr};
     };
 public:
-    // dry run
-    void                    dryRun(const bool isDry);
-    bool                    isDryRun(void) const;
-    // trajectory number
-    void                    setTrajectory(const unsigned int traj);
-    unsigned int            getTrajectory(void) const;
     // grids
     void                    createGrid(const unsigned int Ls);
     GridCartesian *         getGrid(const unsigned int Ls = 1) const;
     GridRedBlackCartesian * getRbGrid(const unsigned int Ls = 1) const;
     std::vector<int>        getDim(void) const;
     int                     getDim(const unsigned int mu) const;
+    unsigned long int       getLocalVolume(void) const;
     unsigned int            getNd(void) const;
     // random number generator
     void                    setSeed(const std::vector<int> &seed);
     GridParallelRNG *       get4dRng(void) const;
-    // module management
-    void                    pushModule(ModPt &pt);
-    template <typename M>
-    void                    createModule(const std::string name);
-    template <typename M>
-    void                    createModule(const std::string name,
-                                         const typename M::Par &par);
-    void                    createModule(const std::string name,
-                                         const std::string type,
-                                         XmlReader &reader);
-    unsigned int            getNModule(void) const;
-    ModuleBase *            getModule(const unsigned int address) const;
-    ModuleBase *            getModule(const std::string name) const;
-    template <typename M>
-    M *                     getModule(const unsigned int address) const;
-    template <typename M>
-    M *                     getModule(const std::string name) const;
-    unsigned int            getModuleAddress(const std::string name) const;
-    std::string             getModuleName(const unsigned int address) const;
-    std::string             getModuleType(const unsigned int address) const;
-    std::string             getModuleType(const std::string name) const;
-    std::string             getModuleNamespace(const unsigned int address) const;
-    std::string             getModuleNamespace(const std::string name) const;
-    bool                    hasModule(const unsigned int address) const;
-    bool                    hasModule(const std::string name) const;
-    Graph<unsigned int>     makeModuleGraph(void) const;
-    Size                    executeProgram(const std::vector<unsigned int> &p);
-    Size                    executeProgram(const std::vector<std::string> &p);
     // general memory management
     void                    addObject(const std::string name,
                                       const int moduleAddress = -1);
-    void                    registerObject(const unsigned int address,
+    template <typename B, typename T, typename ... Ts>
-                                           const unsigned int size,
+    void                    createDerivedObject(const std::string name,
-                                           const unsigned int Ls = 1);
+                                                const Environment::Storage storage,
-    void                    registerObject(const std::string name,
+                                                const unsigned int Ls,
-                                           const unsigned int size,
+                                                Ts && ... args);
-                                           const unsigned int Ls = 1);
+    template <typename T, typename ... Ts>
-    template <typename T>
+    void                    createObject(const std::string name,
-    unsigned int            lattice4dSize(void) const;
+                                         const Environment::Storage storage,
-    template <typename T>
+                                         const unsigned int Ls,
-    void                    registerLattice(const unsigned int address,
+                                         Ts && ... args);
-                                            const unsigned int Ls = 1);
+    void                    setObjectModule(const unsigned int objAddress,
-    template <typename T>
+                                            const int modAddress);
-    void                    registerLattice(const std::string name,
-                                            const unsigned int Ls = 1);
-    template <typename T>
-    void                    setObject(const unsigned int address, T *object);
-    template <typename T>
-    void                    setObject(const std::string name, T *object);
     template <typename T>
     T *                     getObject(const unsigned int address) const;
     template <typename T>
     T *                     getObject(const std::string name) const;
-    template <typename T>
+    unsigned int            getMaxAddress(void) const;
-    T *                     createLattice(const unsigned int address);
-    template <typename T>
-    T *                     createLattice(const std::string name);
     unsigned int            getObjectAddress(const std::string name) const;
     std::string             getObjectName(const unsigned int address) const;
     std::string             getObjectType(const unsigned int address) const;
     std::string             getObjectType(const std::string name) const;
     Size                    getObjectSize(const unsigned int address) const;
     Size                    getObjectSize(const std::string name) const;
-    unsigned int            getObjectModule(const unsigned int address) const;
+    Storage                 getObjectStorage(const unsigned int address) const;
-    unsigned int            getObjectModule(const std::string name) const;
+    Storage                 getObjectStorage(const std::string name) const;
+    int                     getObjectModule(const unsigned int address) const;
+    int                     getObjectModule(const std::string name) const;
     unsigned int            getObjectLs(const unsigned int address) const;
     unsigned int            getObjectLs(const std::string name) const;
     bool                    hasObject(const unsigned int address) const;
     bool                    hasObject(const std::string name) const;
-    bool                    hasRegisteredObject(const unsigned int address) const;
-    bool                    hasRegisteredObject(const std::string name) const;
     bool                    hasCreatedObject(const unsigned int address) const;
     bool                    hasCreatedObject(const std::string name) const;
     bool                    isObject5d(const unsigned int address) const;
@@ -192,20 +138,17 @@ public:
     template <typename T>
     bool                    isObjectOfType(const std::string name) const;
     Environment::Size       getTotalSize(void) const;
-    void                    addOwnership(const unsigned int owner,
+    void                    freeObject(const unsigned int address);
-                                         const unsigned int property);
+    void                    freeObject(const std::string name);
-    void                    addOwnership(const std::string owner,
-                                         const std::string property);
-    bool                    hasOwners(const unsigned int address) const;
-    bool                    hasOwners(const std::string name) const;
-    bool                    freeObject(const unsigned int address);
-    bool                    freeObject(const std::string name);
     void                    freeAll(void);
-    void                    printContent(void);
+    void                    protectObjects(const bool protect);
+    bool                    objectsProtected(void) const;
+    // print environment content
+    void                    printContent(void) const;
 private:
     // general
-    bool                                   dryRun_{false};
+    unsigned long int                      locVol_;
-    unsigned int                           traj_, locVol_;
+    bool                                   protect_{true};
     // grids
     std::vector<int>                       dim_;
     GridPt                                 grid4d_;
@@ -215,11 +158,6 @@ private:
     unsigned int                           nd_;
     // random number generator
     RngPt                                  rng4d_;
-    // module and related maps
-    std::vector<ModuleInfo>                module_;
-    std::map<std::string, unsigned int>    moduleAddress_;
-    // lattice store
-    std::map<unsigned int, LatticePt>      lattice_;
     // object store
     std::vector<ObjInfo>                   object_;
     std::map<std::string, unsigned int>    objectAddress_;
@@ -256,116 +194,85 @@ void Holder<T>::reset(T *pt)
 /******************************************************************************
  *                     Environment template implementation                    *
  ******************************************************************************/
-// module management ///////////////////////////////////////////////////////////
+// general memory management ///////////////////////////////////////////////////
-template <typename M>
+template <typename B, typename T, typename ... Ts>
-void Environment::createModule(const std::string name)
+void Environment::createDerivedObject(const std::string name,
+                                      const Environment::Storage storage,
+                                      const unsigned int Ls,
+                                      Ts && ... args)
 {
-    ModPt pt(new M(name));
+    if (!hasObject(name))
+    {
+        addObject(name);
+    }
     
-    pushModule(pt);
+    unsigned int address = getObjectAddress(name);
-}
-
-template <typename M>
-void Environment::createModule(const std::string name,
-                               const typename M::Par &par)
-{
-    ModPt pt(new M(name));
     
-    static_cast<M *>(pt.get())->setPar(par);
+    if (!object_[address].data or !objectsProtected())
-    pushModule(pt);
-}
-
-template <typename M>
-M * Environment::getModule(const unsigned int address) const
-{
-    if (auto *pt = dynamic_cast<M *>(getModule(address)))
     {
-        return pt;
+        MemoryStats memStats;
+    
+        if (!MemoryProfiler::stats)
+        {
+            MemoryProfiler::stats = &memStats;
+        }
+        size_t initMem           = MemoryProfiler::stats->currentlyAllocated;
+        object_[address].storage = storage;
+        object_[address].Ls      = Ls;
+        object_[address].data.reset(new Holder<B>(new T(std::forward<Ts>(args)...)));
+        object_[address].size    = MemoryProfiler::stats->maxAllocated - initMem;
+        object_[address].type    = &typeid(T);
+        if (MemoryProfiler::stats == &memStats)
+        {
+            MemoryProfiler::stats = nullptr;
+        }
     }
-    else
+    // object already exists, no error if it is a cache, error otherwise
+    else if ((object_[address].storage != Storage::cache) or 
+             (object_[address].storage != storage)        or
+             (object_[address].name    != name)           or
+             (object_[address].type    != &typeid(T)))
     {
-        HADRON_ERROR("module '" + module_[address].name
+        HADRON_ERROR(Definition, "object '" + name + "' already allocated");
-                     + "' does not have type " + typeid(M).name()
-                     + "(object type: " + getModuleType(address) + ")");
     }
 }
 
-template <typename M>
+template <typename T, typename ... Ts>
-M * Environment::getModule(const std::string name) const
+void Environment::createObject(const std::string name, 
+                               const Environment::Storage storage,
+                               const unsigned int Ls,
+                               Ts && ... args)
 {
-    return getModule<M>(getModuleAddress(name));
+    createDerivedObject<T, T>(name, storage, Ls, std::forward<Ts>(args)...);
-}
-
-template <typename T>
-unsigned int Environment::lattice4dSize(void) const
-{
-    return sizeof(typename T::vector_object)/getGrid()->Nsimd();
-}
-
-template <typename T>
-void Environment::registerLattice(const unsigned int address,
-                                  const unsigned int Ls)
-{
-    createGrid(Ls);
-    registerObject(address, Ls*lattice4dSize<T>(), Ls);
-}
-
-template <typename T>
-void Environment::registerLattice(const std::string name, const unsigned int Ls)
-{
-    createGrid(Ls);
-    registerObject(name, Ls*lattice4dSize<T>(), Ls);
-}
-
-template <typename T>
-void Environment::setObject(const unsigned int address, T *object)
-{
-    if (hasRegisteredObject(address))
-    {
-        object_[address].data.reset(new Holder<T>(object));
-        object_[address].type = &typeid(T);
-    }
-    else if (hasObject(address))
-    {
-        HADRON_ERROR("object with address " + std::to_string(address) +
-                     " exists but is not registered");
-    }
-    else
-    {
-        HADRON_ERROR("no object with address " + std::to_string(address));
-    }
-}
-
-template <typename T>
-void Environment::setObject(const std::string name, T *object)
-{
-    setObject(getObjectAddress(name), object);
 }
 
 template <typename T>
 T * Environment::getObject(const unsigned int address) const
 {
-    if (hasRegisteredObject(address))
+    if (hasObject(address))
     {
-        if (auto h = dynamic_cast<Holder<T> *>(object_[address].data.get()))
+        if (hasCreatedObject(address))
         {
-            return h->getPt();
+            if (auto h = dynamic_cast<Holder<T> *>(object_[address].data.get()))
+            {
+                return h->getPt();
+            }
+            else
+            {
+                HADRON_ERROR(Definition, "object with address " + std::to_string(address) +
+                            " does not have type '" + typeName(&typeid(T)) +
+                            "' (has type '" + getObjectType(address) + "')");
+            }
         }
         else
         {
-            HADRON_ERROR("object with address " + std::to_string(address) +
+            HADRON_ERROR(Definition, "object with address " + std::to_string(address) +
-                         " does not have type '" + typeName(&typeid(T)) +
+                         " is empty");
-                         "' (has type '" + getObjectType(address) + "')");
         }
     }
-    else if (hasObject(address))
-    {
-        HADRON_ERROR("object with address " + std::to_string(address) +
-                     " exists but is not registered");
-    }
     else
     {
-        HADRON_ERROR("no object with address " + std::to_string(address));
+        HADRON_ERROR(Definition, "no object with address " + std::to_string(address));
     }
 }
 
@@ -375,26 +282,10 @@ T * Environment::getObject(const std::string name) const
     return getObject<T>(getObjectAddress(name));
 }
 
-template <typename T>
-T * Environment::createLattice(const unsigned int address)
-{
-    GridCartesian *g = getGrid(getObjectLs(address));
-    
-    setObject(address, new T(g));
-    
-    return getObject<T>(address);
-}
-
-template <typename T>
-T * Environment::createLattice(const std::string name)
-{
-    return createLattice<T>(getObjectAddress(name));
-}
-
 template <typename T>
 bool Environment::isObjectOfType(const unsigned int address) const
 {
-    if (hasRegisteredObject(address))
+    if (hasObject(address))
     {
         if (auto h = dynamic_cast<Holder<T> *>(object_[address].data.get()))
         {
@@ -405,14 +296,9 @@ bool Environment::isObjectOfType(const unsigned int address) const
             return false;
         }
     }
-    else if (hasObject(address))
-    {
-        HADRON_ERROR("object with address " + std::to_string(address) +
-                     " exists but is not registered");
-    }
     else
     {
-        HADRON_ERROR("no object with address " + std::to_string(address));
+        HADRON_ERROR(Definition, "no object with address " + std::to_string(address));
     }
 }
 

extras/Hadrons/Exceptions.cc

@@ -0,0 +1,57 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Exceptions.cc
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/Hadrons/Exceptions.hpp>
+
+#ifndef ERR_SUFF
+#define ERR_SUFF " (" + loc + ")"
+#endif
+
+#define CONST_EXC(name, init) \
+name::name(std::string msg, std::string loc)\
+:init\
+{}
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace Exceptions;
+
+// logic errors
+CONST_EXC(Logic, logic_error(msg + ERR_SUFF))
+CONST_EXC(Definition, Logic("definition error: " + msg, loc))
+CONST_EXC(Implementation, Logic("implementation error: " + msg, loc))
+CONST_EXC(Range, Logic("range error: " + msg, loc))
+CONST_EXC(Size, Logic("size error: " + msg, loc))
+// runtime errors
+CONST_EXC(Runtime, runtime_error(msg + ERR_SUFF))
+CONST_EXC(Argument, Runtime("argument error: " + msg, loc))
+CONST_EXC(Io, Runtime("IO error: " + msg, loc))
+CONST_EXC(Memory, Runtime("memory error: " + msg, loc))
+CONST_EXC(Parsing, Runtime("parsing error: " + msg, loc))
+CONST_EXC(Program, Runtime("program error: " + msg, loc))
+CONST_EXC(System, Runtime("system error: " + msg, loc))

extras/Hadrons/Exceptions.hpp

@@ -0,0 +1,72 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Exceptions.hpp
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#ifndef Hadrons_Exceptions_hpp_
+#define Hadrons_Exceptions_hpp_
+
+#include <stdexcept>
+#ifndef Hadrons_Global_hpp_
+#include <Grid/Hadrons/Global.hpp>
+#endif
+
+#define SRC_LOC std::string(__FUNCTION__) + " at " + std::string(__FILE__) + ":"\
+                + std::to_string(__LINE__)
+#define HADRON_ERROR(exc, msg)\
+LOG(Error) << msg << std::endl;\
+throw(Exceptions::exc(msg, SRC_LOC));
+
+#define DECL_EXC(name, base) \
+class name: public base\
+{\
+public:\
+    name(std::string msg, std::string loc);\
+}
+
+BEGIN_HADRONS_NAMESPACE
+
+namespace Exceptions
+{
+    // logic errors
+    DECL_EXC(Logic, std::logic_error);
+    DECL_EXC(Definition, Logic);
+    DECL_EXC(Implementation, Logic);
+    DECL_EXC(Range, Logic);
+    DECL_EXC(Size, Logic);
+    // runtime errors
+    DECL_EXC(Runtime, std::runtime_error);
+    DECL_EXC(Argument, Runtime);
+    DECL_EXC(Io, Runtime);
+    DECL_EXC(Memory, Runtime);
+    DECL_EXC(Parsing, Runtime);
+    DECL_EXC(Program, Runtime);
+    DECL_EXC(System, Runtime);
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Exceptions_hpp_

extras/Hadrons/Factory.hpp

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Factory.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -95,7 +94,7 @@ std::unique_ptr<T> Factory<T>::create(const std::string type,
     }
     catch (std::out_of_range &)
     {
-        HADRON_ERROR("object of type '" + type + "' unknown");
+        HADRON_ERROR(Argument, "object of type '" + type + "' unknown");
     }
     
     return func(name);

extras/Hadrons/GeneticScheduler.hpp

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/GeneticScheduler.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -38,13 +37,13 @@ BEGIN_HADRONS_NAMESPACE
 /******************************************************************************
  *                   Scheduler based on a genetic algorithm                   *
  ******************************************************************************/
-template <typename T>
+template <typename V, typename T>
 class GeneticScheduler
 {
 public:
-    typedef std::vector<T>                   Gene;
+    typedef std::vector<T>                 Gene;
-    typedef std::pair<Gene *, Gene *>        GenePair;
+    typedef std::pair<Gene *, Gene *>      GenePair;
-    typedef std::function<int(const Gene &)> ObjFunc;
+    typedef std::function<V(const Gene &)> ObjFunc;
     struct Parameters
     {
         double       mutationRate;
@@ -65,7 +64,7 @@ public:
     void benchmarkCrossover(const unsigned int nIt);
     // print population
     friend std::ostream & operator<<(std::ostream &out,
-                                     const GeneticScheduler<T> &s)
+                                     const GeneticScheduler<V, T> &s)
     {
         out << "[";
         for (auto &p: s.population_)
@@ -87,19 +86,19 @@ private:
     void     mutation(Gene &m, const Gene &c);
     
 private:
-    Graph<T>                 &graph_;
+    Graph<T>               &graph_;
-    const ObjFunc            &func_;
+    const ObjFunc          &func_;
-    const Parameters         par_;
+    const Parameters       par_;
-    std::multimap<int, Gene> population_;
+    std::multimap<V, Gene> population_;
-    std::mt19937             gen_;
+    std::mt19937           gen_;
 };
 
 /******************************************************************************
  *                       template implementation                              *
  ******************************************************************************/
 // constructor /////////////////////////////////////////////////////////////////
-template <typename T>
+template <typename V, typename T>
-GeneticScheduler<T>::GeneticScheduler(Graph<T> &graph, const ObjFunc &func,
+GeneticScheduler<V, T>::GeneticScheduler(Graph<T> &graph, const ObjFunc &func,
                                       const Parameters &par)
 : graph_(graph)
 , func_(func)
@@ -109,22 +108,22 @@ GeneticScheduler<T>::GeneticScheduler(Graph<T> &graph, const ObjFunc &func,
 }
 
 // access //////////////////////////////////////////////////////////////////////
-template <typename T>
+template <typename V, typename T>
-const typename GeneticScheduler<T>::Gene &
+const typename GeneticScheduler<V, T>::Gene &
-GeneticScheduler<T>::getMinSchedule(void)
+GeneticScheduler<V, T>::getMinSchedule(void)
 {
     return population_.begin()->second;
 }
 
-template <typename T>
+template <typename V, typename T>
-int GeneticScheduler<T>::getMinValue(void)
+int GeneticScheduler<V, T>::getMinValue(void)
 {
     return population_.begin()->first;
 }
 
 // breed a new generation //////////////////////////////////////////////////////
-template <typename T>
+template <typename V, typename T>
-void GeneticScheduler<T>::nextGeneration(void)
+void GeneticScheduler<V, T>::nextGeneration(void)
 {
     // random initialization of the population if necessary
     if (population_.size() != par_.popSize)
@@ -158,8 +157,8 @@ void GeneticScheduler<T>::nextGeneration(void)
 }
 
 // evolution steps /////////////////////////////////////////////////////////////
-template <typename T>
+template <typename V, typename T>
-void GeneticScheduler<T>::initPopulation(void)
+void GeneticScheduler<V, T>::initPopulation(void)
 {
     population_.clear();
     for (unsigned int i = 0; i < par_.popSize; ++i)
@@ -170,8 +169,8 @@ void GeneticScheduler<T>::initPopulation(void)
     }
 }
 
-template <typename T>
+template <typename V, typename T>
-void GeneticScheduler<T>::doCrossover(void)
+void GeneticScheduler<V, T>::doCrossover(void)
 {
     auto p = selectPair();
     Gene &p1 = *(p.first), &p2 = *(p.second);
@@ -185,8 +184,8 @@ void GeneticScheduler<T>::doCrossover(void)
     }
 }
 
-template <typename T>
+template <typename V, typename T>
-void GeneticScheduler<T>::doMutation(void)
+void GeneticScheduler<V, T>::doMutation(void)
 {
     std::uniform_real_distribution<double>      mdis(0., 1.);
     std::uniform_int_distribution<unsigned int> pdis(0, population_.size() - 1);
@@ -206,40 +205,35 @@ void GeneticScheduler<T>::doMutation(void)
 }
 
 // genetic operators ///////////////////////////////////////////////////////////
-template <typename T>
+template <typename V, typename T>
-typename GeneticScheduler<T>::GenePair GeneticScheduler<T>::selectPair(void)
+typename GeneticScheduler<V, T>::GenePair GeneticScheduler<V, T>::selectPair(void)
 {
     std::vector<double> prob;
     unsigned int        ind;
     Gene                *p1, *p2;
+    const double        max = population_.rbegin()->first;
     
+
     for (auto &c: population_)
     {
-        prob.push_back(1./c.first);
+        prob.push_back(std::exp((c.first-1.)/max));
-    }
+    }        
-    do
+    std::discrete_distribution<unsigned int> dis1(prob.begin(), prob.end());
-    {
+    auto rIt = population_.begin();
-        double probCpy;
+    ind = dis1(gen_);
-        
+    std::advance(rIt, ind);
-        std::discrete_distribution<unsigned int> dis1(prob.begin(), prob.end());
+    p1 = &(rIt->second);
-        auto rIt = population_.begin();
+    prob[ind] = 0.;
-        ind = dis1(gen_);
+    std::discrete_distribution<unsigned int> dis2(prob.begin(), prob.end());
-        std::advance(rIt, ind);
+    rIt = population_.begin();
-        p1 = &(rIt->second);
+    std::advance(rIt, dis2(gen_));
-        probCpy   = prob[ind];
+    p2 = &(rIt->second);
-        prob[ind] = 0.;
-        std::discrete_distribution<unsigned int> dis2(prob.begin(), prob.end());
-        rIt = population_.begin();
-        std::advance(rIt, dis2(gen_));
-        p2 = &(rIt->second);
-        prob[ind] = probCpy;
-    } while (p1 == p2);
     
     return std::make_pair(p1, p2);
 }
 
-template <typename T>
+template <typename V, typename T>
-void GeneticScheduler<T>::crossover(Gene &c1, Gene &c2, const Gene &p1,
+void GeneticScheduler<V, T>::crossover(Gene &c1, Gene &c2, const Gene &p1,
                                     const Gene &p2)
 {
     Gene                                        buf;
@@ -273,8 +267,8 @@ void GeneticScheduler<T>::crossover(Gene &c1, Gene &c2, const Gene &p1,
     }
 }
 
-template <typename T>
+template <typename V, typename T>
-void GeneticScheduler<T>::mutation(Gene &m, const Gene &c)
+void GeneticScheduler<V, T>::mutation(Gene &m, const Gene &c)
 {
     Gene                                        buf;
     std::uniform_int_distribution<unsigned int> dis(0, c.size() - 1);
@@ -303,8 +297,8 @@ void GeneticScheduler<T>::mutation(Gene &m, const Gene &c)
     }
 }
 
-template <typename T>
+template <typename V, typename T>
-void GeneticScheduler<T>::benchmarkCrossover(const unsigned int nIt)
+void GeneticScheduler<V, T>::benchmarkCrossover(const unsigned int nIt)
 {
     Gene   p1, p2, c1, c2;
     double neg = 0., eq = 0., pos = 0., total;

extras/Hadrons/Global.cc

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Global.cc
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -39,31 +38,19 @@ HadronsLogger Hadrons::HadronsLogMessage(1,"Message");
 HadronsLogger Hadrons::HadronsLogIterative(1,"Iterative");
 HadronsLogger Hadrons::HadronsLogDebug(1,"Debug");
 
-// pretty size formatting //////////////////////////////////////////////////////
+void Hadrons::initLogger(void)
-std::string Hadrons::sizeString(long unsigned int bytes)
-
 {
-    constexpr unsigned int bufSize = 256;
+    auto w = std::string("Hadrons").length();
-    const char             *suffixes[7] = {"", "K", "M", "G", "T", "P", "E"};
+    GridLogError.setTopWidth(w);
-    char                   buf[256];
+    GridLogWarning.setTopWidth(w);
-    long unsigned int      s     = 0;
+    GridLogMessage.setTopWidth(w);
-    double                 count = bytes;
+    GridLogIterative.setTopWidth(w);
-    
+    GridLogDebug.setTopWidth(w);
-    while (count >= 1024 && s < 7)
+    HadronsLogError.Active(GridLogError.isActive());
-    {
+    HadronsLogWarning.Active(GridLogWarning.isActive());
-        s++;
+    HadronsLogMessage.Active(GridLogMessage.isActive());
-        count /= 1024;
+    HadronsLogIterative.Active(GridLogIterative.isActive());
-    }
+    HadronsLogDebug.Active(GridLogDebug.isActive());
-    if (count - floor(count) == 0.0)
-    {
-        snprintf(buf, bufSize, "%d %sB", (int)count, suffixes[s]);
-    }
-    else
-    {
-        snprintf(buf, bufSize, "%.1f %sB", count, suffixes[s]);
-    }
-    
-    return std::string(buf);
 }
 
 // type utilities //////////////////////////////////////////////////////////////
@@ -80,3 +67,10 @@ std::string Hadrons::typeName(const std::type_info *info)
     
     return name;
 }
+
+// default writers/readers /////////////////////////////////////////////////////
+#ifdef HAVE_HDF5
+const std::string Hadrons::resultFileExt = "h5";
+#else
+const std::string Hadrons::resultFileExt = "xml";
+#endif

extras/Hadrons/Global.hpp

@@ -4,10 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Global.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -35,6 +35,10 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #include <Grid/Grid.h>
 #include <cxxabi.h>
 
+#ifndef SITE_SIZE_TYPE
+#define SITE_SIZE_TYPE size_t
+#endif
+
 #define BEGIN_HADRONS_NAMESPACE \
 namespace Grid {\
 using namespace QCD;\
@@ -57,6 +61,9 @@ using Grid::operator<<;
 #ifndef SIMPL
 #define SIMPL ScalarImplCR
 #endif
+#ifndef GIMPL
+#define GIMPL GimplTypesR
+#endif
 
 BEGIN_HADRONS_NAMESPACE
 
@@ -80,7 +87,8 @@ typedef std::function<void(FermionField##suffix &,\
                       const FermionField##suffix &)> SolverFn##suffix;
 
 #define SINK_TYPE_ALIASES(suffix)\
-typedef std::function<SlicedPropagator##suffix(const PropagatorField##suffix &)> SinkFn##suffix;
+typedef std::function<SlicedPropagator##suffix\
+                      (const PropagatorField##suffix &)> SinkFn##suffix;
 
 #define FGS_TYPE_ALIASES(FImpl, suffix)\
 FERM_TYPE_ALIASES(FImpl, suffix)\
@@ -96,11 +104,6 @@ public:
 };
 
 #define LOG(channel) std::cout << HadronsLog##channel
-#define HADRON_ERROR(msg)\
-LOG(Error) << msg << " (" << __FUNCTION__ << " at " << __FILE__ << ":"\
-           << __LINE__ << ")" << std::endl;\
-abort();
-
 #define DEBUG_VAR(var) LOG(Debug) << #var << "= " << (var) << std::endl;
 
 extern HadronsLogger HadronsLogError;
@@ -109,6 +112,8 @@ extern HadronsLogger HadronsLogMessage;
 extern HadronsLogger HadronsLogIterative;
 extern HadronsLogger HadronsLogDebug;
 
+void initLogger(void);
+
 // singleton pattern
 #define SINGLETON(name)\
 public:\
@@ -134,9 +139,6 @@ public:\
 private:\
     name(void) = default;
 
-// pretty size formating
-std::string sizeString(long unsigned int bytes);
-
 // type utilities
 template <typename T>
 const std::type_info * typeIdPt(const T &x)
@@ -165,14 +167,21 @@ std::string typeName(void)
 }
 
 // default writers/readers
+extern const std::string resultFileExt;
+
 #ifdef HAVE_HDF5
-typedef Hdf5Reader CorrReader;
+typedef Hdf5Reader ResultReader;
-typedef Hdf5Writer CorrWriter;
+typedef Hdf5Writer ResultWriter;
 #else
-typedef XmlReader CorrReader;
+typedef XmlReader ResultReader;
-typedef XmlWriter CorrWriter;
+typedef XmlWriter ResultWriter;
 #endif
 
+#define RESULT_FILE_NAME(name) \
+name + "." + std::to_string(vm().getTrajectory()) + "." + resultFileExt
+
 END_HADRONS_NAMESPACE
 
+#include <Grid/Hadrons/Exceptions.hpp>
+
 #endif // Hadrons_Global_hpp_

extras/Hadrons/Graph.hpp

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Graph.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -185,7 +184,7 @@ void Graph<T>::removeVertex(const T &value)
     }
     else
     {
-        HADRON_ERROR("vertex " << value << " does not exists");
+        HADRON_ERROR(Range, "vertex does not exists");
     }
 
     // remove all edges containing the vertex
@@ -214,7 +213,7 @@ void Graph<T>::removeEdge(const Edge &e)
     }
     else
     {
-        HADRON_ERROR("edge "  << e << " does not exists");
+        HADRON_ERROR(Range, "edge does not exists");
     }
 }
 
@@ -260,7 +259,7 @@ void Graph<T>::mark(const T &value, const bool doMark)
     }
     else
     {
-        HADRON_ERROR("vertex " << value << " does not exists");
+        HADRON_ERROR(Range, "vertex does not exists");
     }
 }
 
@@ -298,7 +297,7 @@ bool Graph<T>::isMarked(const T &value) const
     }
     else
     {
-        HADRON_ERROR("vertex " << value << " does not exists");
+        HADRON_ERROR(Range, "vertex does not exists");
         
         return false;
     }
@@ -430,7 +429,7 @@ std::vector<T> Graph<T>::getAdjacentVertices(const T &value) const
     {
         return ((e.first == value) or (e.second == value));
     };
-    auto eIt = find_if(edgeSet_.begin(), edgeSet_.end(), pred);
+    auto eIt = std::find_if(edgeSet_.begin(), edgeSet_.end(), pred);
     
     while (eIt != edgeSet_.end())
     {
@@ -442,7 +441,7 @@ std::vector<T> Graph<T>::getAdjacentVertices(const T &value) const
         {
             adjacentVertex.push_back((*eIt).first);
         }
-        eIt = find_if(++eIt, edgeSet_.end(), pred);
+        eIt = std::find_if(++eIt, edgeSet_.end(), pred);
     }
     
     return adjacentVertex;
@@ -458,12 +457,12 @@ std::vector<T> Graph<T>::getChildren(const T &value) const
     {
         return (e.first == value);
     };
-    auto eIt = find_if(edgeSet_.begin(), edgeSet_.end(), pred);
+    auto eIt = std::find_if(edgeSet_.begin(), edgeSet_.end(), pred);
     
     while (eIt != edgeSet_.end())
     {
         child.push_back((*eIt).second);
-        eIt = find_if(++eIt, edgeSet_.end(), pred);
+        eIt = std::find_if(++eIt, edgeSet_.end(), pred);
     }
     
     return child;
@@ -479,12 +478,12 @@ std::vector<T> Graph<T>::getParents(const T &value) const
     {
         return (e.second == value);
     };
-    auto eIt = find_if(edgeSet_.begin(), edgeSet_.end(), pred);
+    auto eIt = std::find_if(edgeSet_.begin(), edgeSet_.end(), pred);
     
     while (eIt != edgeSet_.end())
     {
         parent.push_back((*eIt).first);
-        eIt = find_if(++eIt, edgeSet_.end(), pred);
+        eIt = std::find_if(++eIt, edgeSet_.end(), pred);
     }
     
     return parent;
@@ -544,7 +543,7 @@ std::vector<T> Graph<T>::topoSort(void)
     {
         if (tmpMarked.at(v))
         {
-            HADRON_ERROR("cannot topologically sort a cyclic graph");
+            HADRON_ERROR(Range, "cannot topologically sort a cyclic graph");
         }
         if (!isMarked(v))
         {
@@ -603,7 +602,7 @@ std::vector<T> Graph<T>::topoSort(Gen &gen)
     {
         if (tmpMarked.at(v))
         {
-            HADRON_ERROR("cannot topologically sort a cyclic graph");
+            HADRON_ERROR(Range, "cannot topologically sort a cyclic graph");
         }
         if (!isMarked(v))
         {

extras/Hadrons/HadronsXmlRun.cc

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/HadronsXmlRun.cc
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -55,12 +54,6 @@ int main(int argc, char *argv[])
     
     // initialization
     Grid_init(&argc, &argv);
-    HadronsLogError.Active(GridLogError.isActive());
-    HadronsLogWarning.Active(GridLogWarning.isActive());
-    HadronsLogMessage.Active(GridLogMessage.isActive());
-    HadronsLogIterative.Active(GridLogIterative.isActive());
-    HadronsLogDebug.Active(GridLogDebug.isActive());
-    LOG(Message) << "Grid initialized" << std::endl;
     
     // execution
     Application application(parameterFileName);

extras/Hadrons/HadronsXmlSchedule.cc

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/HadronsXmlSchedule.cc
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -49,12 +48,6 @@ int main(int argc, char *argv[])
     
     // initialization
     Grid_init(&argc, &argv);
-    HadronsLogError.Active(GridLogError.isActive());
-    HadronsLogWarning.Active(GridLogWarning.isActive());
-    HadronsLogMessage.Active(GridLogMessage.isActive());
-    HadronsLogIterative.Active(GridLogIterative.isActive());
-    HadronsLogDebug.Active(GridLogDebug.isActive());
-    LOG(Message) << "Grid initialized" << std::endl;
     
     // execution
     Application application;

extras/Hadrons/Makefile.am

@@ -7,20 +7,24 @@ libHadrons_a_SOURCES = \
     $(modules_cc)      \
     Application.cc     \
     Environment.cc     \
+	Exceptions.cc      \
     Global.cc          \
-    Module.cc
+    Module.cc		   \
+	VirtualMachine.cc
 libHadrons_adir = $(pkgincludedir)/Hadrons
 nobase_libHadrons_a_HEADERS = \
 	$(modules_hpp)            \
 	Application.hpp           \
 	Environment.hpp           \
+	Exceptions.hpp            \
 	Factory.hpp               \
 	GeneticScheduler.hpp      \
 	Global.hpp                \
 	Graph.hpp                 \
 	Module.hpp                \
 	Modules.hpp               \
-	ModuleFactory.hpp
+	ModuleFactory.hpp         \
+	VirtualMachine.hpp
 
 HadronsXmlRun_SOURCES = HadronsXmlRun.cc
 HadronsXmlRun_LDADD   = libHadrons.a -lGrid

extras/Hadrons/Module.cc

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Module.cc
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -39,7 +38,6 @@ using namespace Hadrons;
 // constructor /////////////////////////////////////////////////////////////////
 ModuleBase::ModuleBase(const std::string name)
 : name_(name)
-, env_(Environment::getInstance())
 {}
 
 // access //////////////////////////////////////////////////////////////////////
@@ -48,15 +46,10 @@ std::string ModuleBase::getName(void) const
     return name_;
 }
 
-Environment & ModuleBase::env(void) const
-{
-    return env_;
-}
-
 // get factory registration name if available
 std::string ModuleBase::getRegisteredName(void)
 {
-    HADRON_ERROR("module '" + getName() + "' has a type not registered"
+    HADRON_ERROR(Definition, "module '" + getName() + "' has no registered type"
                  + " in the factory");
 }
 
@@ -64,8 +57,5 @@ std::string ModuleBase::getRegisteredName(void)
 void ModuleBase::operator()(void)
 {
     setup();
-    if (!env().isDryRun())
+    execute();
-    {
-        execute();
-    }
 }

extras/Hadrons/Module.hpp

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Module.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -31,7 +30,7 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #define Hadrons_Module_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
-#include <Grid/Hadrons/Environment.hpp>
+#include <Grid/Hadrons/VirtualMachine.hpp>
 
 BEGIN_HADRONS_NAMESPACE
 
@@ -87,6 +86,56 @@ public:\
 static ns##mod##ModuleRegistrar ns##mod##ModuleRegistrarInstance;
 
 #define ARG(...) __VA_ARGS__
+#define MACRO_REDIRECT(arg1, arg2, arg3, macro, ...) macro
+
+#define envGet(type, name)\
+*env().template getObject<type>(name)
+
+#define envGetTmp(type, var)\
+type &var = *env().template getObject<type>(getName() + "_tmp_" + #var)
+
+#define envHasType(type, name)\
+env().template isObjectOfType<type>(name)
+
+#define envCreate(type, name, Ls, ...)\
+env().template createObject<type>(name, Environment::Storage::object, Ls, __VA_ARGS__)
+
+#define envCreateDerived(base, type, name, Ls, ...)\
+env().template createDerivedObject<base, type>(name, Environment::Storage::object, Ls, __VA_ARGS__)
+
+#define envCreateLat4(type, name)\
+envCreate(type, name, 1, env().getGrid())
+
+#define envCreateLat5(type, name, Ls)\
+envCreate(type, name, Ls, env().getGrid(Ls))
+
+#define envCreateLat(...)\
+MACRO_REDIRECT(__VA_ARGS__, envCreateLat5, envCreateLat4)(__VA_ARGS__)
+
+#define envCache(type, name, Ls, ...)\
+env().template createObject<type>(name, Environment::Storage::cache, Ls, __VA_ARGS__)
+
+#define envCacheLat4(type, name)\
+envCache(type, name, 1, env().getGrid())
+
+#define envCacheLat5(type, name, Ls)\
+envCache(type, name, Ls, env().getGrid(Ls))
+
+#define envCacheLat(...)\
+MACRO_REDIRECT(__VA_ARGS__, envCacheLat5, envCacheLat4)(__VA_ARGS__)
+
+#define envTmp(type, name, Ls, ...)\
+env().template createObject<type>(getName() + "_tmp_" + name,         \
+                                  Environment::Storage::temporary, Ls, __VA_ARGS__)
+
+#define envTmpLat4(type, name)\
+envTmp(type, name, 1, env().getGrid())
+
+#define envTmpLat5(type, name, Ls)\
+envTmp(type, name, Ls, env().getGrid(Ls))
+
+#define envTmpLat(...)\
+MACRO_REDIRECT(__VA_ARGS__, envTmpLat5, envTmpLat4)(__VA_ARGS__)
 
 /******************************************************************************
  *                            Module class                                    *
@@ -101,23 +150,30 @@ public:
     virtual ~ModuleBase(void) = default;
     // access
     std::string getName(void) const;
-    Environment &env(void) const;
     // get factory registration name if available
     virtual std::string getRegisteredName(void);
     // dependencies/products
     virtual std::vector<std::string> getInput(void) = 0;
+    virtual std::vector<std::string> getReference(void)
+    {
+        return std::vector<std::string>(0);
+    };
     virtual std::vector<std::string> getOutput(void) = 0;
     // parse parameters
     virtual void parseParameters(XmlReader &reader, const std::string name) = 0;
     virtual void saveParameters(XmlWriter &writer, const std::string name) = 0;
     // setup
     virtual void setup(void) {};
+    virtual void execute(void) = 0;
     // execution
     void operator()(void);
-    virtual void execute(void) = 0;
+protected:
+    // environment shortcut
+    DEFINE_ENV_ALIAS;
+    // virtual machine shortcut
+    DEFINE_VM_ALIAS;
 private:
     std::string name_;
-    Environment &env_;
 };
 
 // derived class, templating the parameter class

extras/Hadrons/ModuleFactory.hpp

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/ModuleFactory.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 

extras/Hadrons/Modules.hpp

@@ -1,31 +1,55 @@
-#include <Grid/Hadrons/Modules/MSolver/RBPrecCG.hpp>
+/*************************************************************************************
-#include <Grid/Hadrons/Modules/MContraction/WardIdentity.hpp>
+Grid physics library, www.github.com/paboyle/Grid 
-#include <Grid/Hadrons/Modules/MContraction/Meson.hpp>
+Source file: extras/Hadrons/Modules.hpp
-#include <Grid/Hadrons/Modules/MContraction/Gamma3pt.hpp>
+Copyright (C) 2015-2018
-#include <Grid/Hadrons/Modules/MContraction/DiscLoop.hpp>
+Author: Antonin Portelli <antonin.portelli@me.com>
-#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp>
+Author: Lanny91 <andrew.lawson@gmail.com>
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
 #include <Grid/Hadrons/Modules/MContraction/Baryon.hpp>
+#include <Grid/Hadrons/Modules/MContraction/Meson.hpp>
 #include <Grid/Hadrons/Modules/MContraction/WeakHamiltonian.hpp>
-#include <Grid/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp>
 #include <Grid/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.hpp>
+#include <Grid/Hadrons/Modules/MContraction/DiscLoop.hpp>
+#include <Grid/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp>
+#include <Grid/Hadrons/Modules/MContraction/Gamma3pt.hpp>
+#include <Grid/Hadrons/Modules/MContraction/WardIdentity.hpp>
+#include <Grid/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp>
+#include <Grid/Hadrons/Modules/MFermion/GaugeProp.hpp>
+#include <Grid/Hadrons/Modules/MSource/SeqConservedSummed.hpp>
+#include <Grid/Hadrons/Modules/MSource/SeqGamma.hpp>
+#include <Grid/Hadrons/Modules/MSource/Point.hpp>
+#include <Grid/Hadrons/Modules/MSource/Wall.hpp>
+#include <Grid/Hadrons/Modules/MSource/Z2.hpp>
+#include <Grid/Hadrons/Modules/MSource/SeqConserved.hpp>
+#include <Grid/Hadrons/Modules/MSink/Smear.hpp>
+#include <Grid/Hadrons/Modules/MSink/Point.hpp>
+#include <Grid/Hadrons/Modules/MSolver/RBPrecCG.hpp>
+#include <Grid/Hadrons/Modules/MGauge/Unit.hpp>
+#include <Grid/Hadrons/Modules/MGauge/Random.hpp>
+#include <Grid/Hadrons/Modules/MGauge/StochEm.hpp>
 #include <Grid/Hadrons/Modules/MUtilities/TestSeqGamma.hpp>
 #include <Grid/Hadrons/Modules/MUtilities/TestSeqConserved.hpp>
 #include <Grid/Hadrons/Modules/MLoop/NoiseLoop.hpp>
 #include <Grid/Hadrons/Modules/MScalar/FreeProp.hpp>
-#include <Grid/Hadrons/Modules/MScalar/ChargedProp.hpp>
 #include <Grid/Hadrons/Modules/MScalar/Scalar.hpp>
-#include <Grid/Hadrons/Modules/MSink/Point.hpp>
+#include <Grid/Hadrons/Modules/MScalar/ChargedProp.hpp>
-#include <Grid/Hadrons/Modules/MSink/Smear.hpp>
-#include <Grid/Hadrons/Modules/MFermion/GaugeProp.hpp>
-#include <Grid/Hadrons/Modules/MSource/SeqConservedSummed.hpp>
-#include <Grid/Hadrons/Modules/MSource/Wall.hpp>
-#include <Grid/Hadrons/Modules/MSource/Point.hpp>
-#include <Grid/Hadrons/Modules/MSource/SeqConserved.hpp>
-#include <Grid/Hadrons/Modules/MSource/SeqGamma.hpp>
-#include <Grid/Hadrons/Modules/MSource/Z2.hpp>
-#include <Grid/Hadrons/Modules/MGauge/Load.hpp>
-#include <Grid/Hadrons/Modules/MGauge/StochEm.hpp>
-#include <Grid/Hadrons/Modules/MGauge/Random.hpp>
-#include <Grid/Hadrons/Modules/MGauge/Unit.hpp>
 #include <Grid/Hadrons/Modules/MAction/DWF.hpp>
 #include <Grid/Hadrons/Modules/MAction/Wilson.hpp>
+#include <Grid/Hadrons/Modules/MScalarSUN/TrMag.hpp>
+#include <Grid/Hadrons/Modules/MScalarSUN/TwoPoint.hpp>
+#include <Grid/Hadrons/Modules/MScalarSUN/TrPhi.hpp>
+#include <Grid/Hadrons/Modules/MIO/LoadNersc.hpp>
+#include <Grid/Hadrons/Modules/MIO/LoadBinary.hpp>

extras/Hadrons/Modules/MAction/DWF.hpp

@@ -4,10 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MAction/DWF.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -65,6 +65,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -103,35 +104,29 @@ std::vector<std::string> TDWF<FImpl>::getOutput(void)
 template <typename FImpl>
 void TDWF<FImpl>::setup(void)
 {
-    unsigned int size;
+    LOG(Message) << "Setting up domain wall fermion matrix with m= "
-    
+                 << par().mass << ", M5= " << par().M5 << " and Ls= "
-    size = 2*env().template lattice4dSize<typename FImpl::DoubledGaugeField>();
+                 << par().Ls << " using gauge field '" << par().gauge << "'"
-    env().registerObject(getName(), size, par().Ls);
+                 << std::endl;
+    LOG(Message) << "Fermion boundary conditions: " << par().boundary
+                 << std::endl;
+                 
+    env().createGrid(par().Ls);
+    auto &U    = envGet(LatticeGaugeField, par().gauge);
+    auto &g4   = *env().getGrid();
+    auto &grb4 = *env().getRbGrid();
+    auto &g5   = *env().getGrid(par().Ls);
+    auto &grb5 = *env().getRbGrid(par().Ls);
+    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
+    typename DomainWallFermion<FImpl>::ImplParams implParams(boundary);
+    envCreateDerived(FMat, DomainWallFermion<FImpl>, getName(), par().Ls, U, g5,
+                     grb5, g4, grb4, par().mass, par().M5, implParams);
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 template <typename FImpl>
 void TDWF<FImpl>::execute(void)
-{
+{}
-    LOG(Message) << "Setting up domain wall fermion matrix with m= "
-                 << par().mass << ", M5= " << par().M5 << " and Ls= "
-                 << par().Ls << " using gauge field '" << par().gauge << "'"
-                 << std::endl;
-    LOG(Message) << "Fermion boundary conditions: " << par().boundary 
-                 << std::endl;
-    env().createGrid(par().Ls);
-    auto &U      = *env().template getObject<LatticeGaugeField>(par().gauge);
-    auto &g4     = *env().getGrid();
-    auto &grb4   = *env().getRbGrid();
-    auto &g5     = *env().getGrid(par().Ls);
-    auto &grb5   = *env().getRbGrid(par().Ls);
-    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
-    typename DomainWallFermion<FImpl>::ImplParams implParams(boundary);
-    FMat *fMatPt = new DomainWallFermion<FImpl>(U, g5, grb5, g4, grb4,
-                                                par().mass, par().M5,
-                                                implParams);
-    env().setObject(getName(), fMatPt);
-}
 
 END_MODULE_NAMESPACE
 

extras/Hadrons/Modules/MAction/Wilson.hpp

@@ -4,10 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MAction/Wilson.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -63,6 +63,7 @@ public:
     // dependencies/products
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -101,29 +102,24 @@ std::vector<std::string> TWilson<FImpl>::getOutput(void)
 template <typename FImpl>
 void TWilson<FImpl>::setup(void)
 {
-    unsigned int size;
+    LOG(Message) << "Setting up TWilson fermion matrix with m= " << par().mass
-    
+                 << " using gauge field '" << par().gauge << "'" << std::endl;
-    size = 2*env().template lattice4dSize<typename FImpl::DoubledGaugeField>();
+    LOG(Message) << "Fermion boundary conditions: " << par().boundary
-    env().registerObject(getName(), size);
+                 << std::endl;
+                 
+    auto &U      = envGet(LatticeGaugeField, par().gauge);
+    auto &grid   = *env().getGrid();
+    auto &gridRb = *env().getRbGrid();
+    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
+    typename WilsonFermion<FImpl>::ImplParams implParams(boundary);
+    envCreateDerived(FMat, WilsonFermion<FImpl>, getName(), 1, U, grid, gridRb,
+                     par().mass, implParams);
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 template <typename FImpl>
 void TWilson<FImpl>::execute()
-{
+{}
-    LOG(Message) << "Setting up TWilson fermion matrix with m= " << par().mass
-                 << " using gauge field '" << par().gauge << "'" << std::endl;
-    LOG(Message) << "Fermion boundary conditions: " << par().boundary 
-                 << std::endl;
-    auto &U      = *env().template getObject<LatticeGaugeField>(par().gauge);
-    auto &grid   = *env().getGrid();
-    auto &gridRb = *env().getRbGrid();
-    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
-    typename WilsonFermion<FImpl>::ImplParams implParams(boundary);
-    FMat *fMatPt = new WilsonFermion<FImpl>(U, grid, gridRb, par().mass,
-                                            implParams);
-    env().setObject(getName(), fMatPt);
-}
 
 END_MODULE_NAMESPACE
 

extras/Hadrons/Modules/MContraction/Baryon.hpp

@@ -4,10 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/Baryon.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -72,6 +72,9 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
+    // setup
+    virtual void setup(void);
     // execution
     virtual void execute(void);
 };
@@ -99,11 +102,18 @@ std::vector<std::string> TBaryon<FImpl1, FImpl2, FImpl3>::getInput(void)
 template <typename FImpl1, typename FImpl2, typename FImpl3>
 std::vector<std::string> TBaryon<FImpl1, FImpl2, FImpl3>::getOutput(void)
 {
-    std::vector<std::string> out = {getName()};
+    std::vector<std::string> out = {};
     
     return out;
 }
 
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2, typename FImpl3>
+void TBaryon<FImpl1, FImpl2, FImpl3>::setup(void)
+{
+    envTmpLat(LatticeComplex, "c");
+}
+
 // execution ///////////////////////////////////////////////////////////////////
 template <typename FImpl1, typename FImpl2, typename FImpl3>
 void TBaryon<FImpl1, FImpl2, FImpl3>::execute(void)
@@ -112,12 +122,12 @@ void TBaryon<FImpl1, FImpl2, FImpl3>::execute(void)
                  << " quarks '" << par().q1 << "', '" << par().q2 << "', and '"
                  << par().q3 << "'" << std::endl;
     
-    CorrWriter             writer(par().output);
+    ResultWriter writer(RESULT_FILE_NAME(par().output));
-    PropagatorField1      &q1 = *env().template getObject<PropagatorField1>(par().q1);
+    auto       &q1 = envGet(PropagatorField1, par().q1);
-    PropagatorField2      &q2 = *env().template getObject<PropagatorField2>(par().q2);
+    auto       &q2 = envGet(PropagatorField2, par().q2);
-    PropagatorField3      &q3 = *env().template getObject<PropagatorField3>(par().q2);
+    auto       &q3 = envGet(PropagatorField3, par().q2);
-    LatticeComplex        c(env().getGrid());
+    envGetTmp(LatticeComplex, c);
-    Result                result;
+    Result     result;
     
     // FIXME: do contractions
     

extras/Hadrons/Modules/MContraction/DiscLoop.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/DiscLoop.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -68,6 +69,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -97,7 +99,7 @@ std::vector<std::string> TDiscLoop<FImpl>::getInput(void)
 template <typename FImpl>
 std::vector<std::string> TDiscLoop<FImpl>::getOutput(void)
 {
-    std::vector<std::string> out = {getName()};
+    std::vector<std::string> out = {};
     
     return out;
 }
@@ -106,7 +108,7 @@ std::vector<std::string> TDiscLoop<FImpl>::getOutput(void)
 template <typename FImpl>
 void TDiscLoop<FImpl>::setup(void)
 {
-    
+    envTmpLat(LatticeComplex, "c");
 }
 
 // execution ///////////////////////////////////////////////////////////////////
@@ -117,13 +119,13 @@ void TDiscLoop<FImpl>::execute(void)
                  << "' using '" << par().q_loop << "' with " << par().gamma 
                  << " insertion." << std::endl;
 
-    CorrWriter            writer(par().output);
+    ResultWriter          writer(RESULT_FILE_NAME(par().output));
-    PropagatorField       &q_loop = *env().template getObject<PropagatorField>(par().q_loop);
+    auto                  &q_loop = envGet(PropagatorField, par().q_loop);
-    LatticeComplex        c(env().getGrid());
     Gamma                 gamma(par().gamma);
     std::vector<TComplex> buf;
     Result                result;
 
+    envGetTmp(LatticeComplex, c);
     c = trace(gamma*q_loop);
     sliceSum(c, buf, Tp);
 

extras/Hadrons/Modules/MContraction/Gamma3pt.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/Gamma3pt.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -99,6 +100,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -128,7 +130,7 @@ std::vector<std::string> TGamma3pt<FImpl1, FImpl2, FImpl3>::getInput(void)
 template <typename FImpl1, typename FImpl2, typename FImpl3>
 std::vector<std::string> TGamma3pt<FImpl1, FImpl2, FImpl3>::getOutput(void)
 {
-    std::vector<std::string> out = {getName()};
+    std::vector<std::string> out = {};
     
     return out;
 }
@@ -137,7 +139,7 @@ std::vector<std::string> TGamma3pt<FImpl1, FImpl2, FImpl3>::getOutput(void)
 template <typename FImpl1, typename FImpl2, typename FImpl3>
 void TGamma3pt<FImpl1, FImpl2, FImpl3>::setup(void)
 {
-    
+    envTmpLat(LatticeComplex, "c");
 }
 
 // execution ///////////////////////////////////////////////////////////////////
@@ -151,11 +153,10 @@ void TGamma3pt<FImpl1, FImpl2, FImpl3>::execute(void)
 
     // Initialise variables. q2 and q3 are normal propagators, q1 may be 
     // sink smeared.
-    CorrWriter            writer(par().output);
+    ResultWriter          writer(RESULT_FILE_NAME(par().output));
-    SlicedPropagator1     &q1 = *env().template getObject<SlicedPropagator1>(par().q1);
+    auto                  &q1 = envGet(SlicedPropagator1, par().q1);
-    PropagatorField2      &q2 = *env().template getObject<PropagatorField2>(par().q2);
+    auto                  &q2 = envGet(PropagatorField2, par().q2);
-    PropagatorField3      &q3 = *env().template getObject<PropagatorField2>(par().q3);
+    auto                  &q3 = envGet(PropagatorField2, par().q3);
-    LatticeComplex        c(env().getGrid());
     Gamma                 g5(Gamma::Algebra::Gamma5);
     Gamma                 gamma(par().gamma);
     std::vector<TComplex> buf;
@@ -164,6 +165,7 @@ void TGamma3pt<FImpl1, FImpl2, FImpl3>::execute(void)
     // Extract relevant timeslice of sinked propagator q1, then contract &
     // sum over all spacial positions of gamma insertion.
     SitePropagator1 q1Snk = q1[par().tSnk];
+    envGetTmp(LatticeComplex, c);
     c = trace(g5*q1Snk*adj(q2)*(g5*gamma)*q3);
     sliceSum(c, buf, Tp);
 

extras/Hadrons/Modules/MContraction/Meson.hpp

@@ -4,12 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/Meson.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
-Copyright (C) 2017
 
 Author: Antonin Portelli <antonin.portelli@me.com>
-        Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -97,6 +95,9 @@ public:
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
     virtual void parseGammaString(std::vector<GammaPair> &gammaList);
+protected:
+    // execution
+    virtual void setup(void);
     // execution
     virtual void execute(void);
 };
@@ -124,7 +125,7 @@ std::vector<std::string> TMeson<FImpl1, FImpl2>::getInput(void)
 template <typename FImpl1, typename FImpl2>
 std::vector<std::string> TMeson<FImpl1, FImpl2>::getOutput(void)
 {
-    std::vector<std::string> output = {getName()};
+    std::vector<std::string> output = {};
     
     return output;
 }
@@ -150,9 +151,15 @@ void TMeson<FImpl1, FImpl2>::parseGammaString(std::vector<GammaPair> &gammaList)
     {
         // Parse individual contractions from input string.
         gammaList = strToVec<GammaPair>(par().gammas);
-    }
+    } 
 }
 
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl1, typename FImpl2>
+void TMeson<FImpl1, FImpl2>::setup(void)
+{
+    envTmpLat(LatticeComplex, "c");
+}
 
 // execution ///////////////////////////////////////////////////////////////////
 #define mesonConnected(q1, q2, gSnk, gSrc) \
@@ -165,7 +172,7 @@ void TMeson<FImpl1, FImpl2>::execute(void)
                  << " quarks '" << par().q1 << "' and '" << par().q2 << "'"
                  << std::endl;
     
-    CorrWriter             writer(par().output);
+    ResultWriter           writer(RESULT_FILE_NAME(par().output));
     std::vector<TComplex>  buf;
     std::vector<Result>    result;
     Gamma                  g5(Gamma::Algebra::Gamma5);
@@ -180,11 +187,11 @@ void TMeson<FImpl1, FImpl2>::execute(void)
         result[i].gamma_src = gammaList[i].second;
         result[i].corr.resize(nt);
     }
-    if (env().template isObjectOfType<SlicedPropagator1>(par().q1) and
+    if (envHasType(SlicedPropagator1, par().q1) and
-        env().template isObjectOfType<SlicedPropagator2>(par().q2))
+        envHasType(SlicedPropagator2, par().q2))
     {
-        SlicedPropagator1 &q1 = *env().template getObject<SlicedPropagator1>(par().q1);
+        auto &q1 = envGet(SlicedPropagator1, par().q1);
-        SlicedPropagator2 &q2 = *env().template getObject<SlicedPropagator2>(par().q2);
+        auto &q2 = envGet(SlicedPropagator2, par().q2);
         
         LOG(Message) << "(propagator already sinked)" << std::endl;
         for (unsigned int i = 0; i < result.size(); ++i)
@@ -200,10 +207,10 @@ void TMeson<FImpl1, FImpl2>::execute(void)
     }
     else
     {
-        PropagatorField1 &q1   = *env().template getObject<PropagatorField1>(par().q1);
+        auto &q1 = envGet(PropagatorField1, par().q1);
-        PropagatorField2 &q2   = *env().template getObject<PropagatorField2>(par().q2);
+        auto &q2 = envGet(PropagatorField2, par().q2);
-        LatticeComplex   c(env().getGrid());
         
+        envGetTmp(LatticeComplex, c);
         LOG(Message) << "(using sink '" << par().sink << "')" << std::endl;
         for (unsigned int i = 0; i < result.size(); ++i)
         {
@@ -211,18 +218,17 @@ void TMeson<FImpl1, FImpl2>::execute(void)
             Gamma       gSrc(gammaList[i].second);
             std::string ns;
                 
-            ns = env().getModuleNamespace(env().getObjectModule(par().sink));
+            ns = vm().getModuleNamespace(env().getObjectModule(par().sink));
             if (ns == "MSource")
             {
-                PropagatorField1 &sink =
+                PropagatorField1 &sink = envGet(PropagatorField1, par().sink);
-                    *env().template getObject<PropagatorField1>(par().sink);
                 
                 c = trace(mesonConnected(q1, q2, gSnk, gSrc)*sink);
                 sliceSum(c, buf, Tp);
             }
             else if (ns == "MSink")
             {
-                SinkFnScalar &sink = *env().template getObject<SinkFnScalar>(par().sink);
+                SinkFnScalar &sink = envGet(SinkFnScalar, par().sink);
                 
                 c   = trace(mesonConnected(q1, q2, gSnk, gSrc));
                 buf = sink(c);

extras/Hadrons/Modules/MContraction/WardIdentity.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/WardIdentity.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -74,6 +75,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -105,7 +107,7 @@ std::vector<std::string> TWardIdentity<FImpl>::getInput(void)
 template <typename FImpl>
 std::vector<std::string> TWardIdentity<FImpl>::getOutput(void)
 {
-    std::vector<std::string> out = {getName()};
+    std::vector<std::string> out = {};
     
     return out;
 }
@@ -117,7 +119,16 @@ void TWardIdentity<FImpl>::setup(void)
     Ls_ = env().getObjectLs(par().q);
     if (Ls_ != env().getObjectLs(par().action))
     {
-        HADRON_ERROR("Ls mismatch between quark action and propagator");
+        HADRON_ERROR(Size, "Ls mismatch between quark action and propagator");
+    }
+    envTmpLat(PropagatorField, "tmp");
+    envTmpLat(PropagatorField, "vector_WI");
+    if (par().test_axial)
+    {
+        envTmpLat(PropagatorField, "psi");
+        envTmpLat(LatticeComplex,  "PP");
+        envTmpLat(LatticeComplex,  "axial_defect");
+        envTmpLat(LatticeComplex,  "PJ5q");
     }
 }
 
@@ -128,12 +139,13 @@ void TWardIdentity<FImpl>::execute(void)
     LOG(Message) << "Performing Ward Identity checks for quark '" << par().q
                  << "'." << std::endl;
 
-    PropagatorField tmp(env().getGrid()), vector_WI(env().getGrid());
+    auto  &q   = envGet(PropagatorField, par().q);
-    PropagatorField &q    = *env().template getObject<PropagatorField>(par().q);
+    auto  &act = envGet(FMat, par().action);
-    FMat            &act = *(env().template getObject<FMat>(par().action));
+    Gamma g5(Gamma::Algebra::Gamma5);
-    Gamma           g5(Gamma::Algebra::Gamma5);
 
     // Compute D_mu V_mu, D here is backward derivative.
+    envGetTmp(PropagatorField, tmp);
+    envGetTmp(PropagatorField, vector_WI);
     vector_WI    = zero;
     for (unsigned int mu = 0; mu < Nd; ++mu)
     {
@@ -148,9 +160,10 @@ void TWardIdentity<FImpl>::execute(void)
 
     if (par().test_axial)
     {
-        PropagatorField psi(env().getGrid());
+        envGetTmp(PropagatorField, psi);
-        LatticeComplex PP(env().getGrid()), axial_defect(env().getGrid()),
+        envGetTmp(LatticeComplex, PP);
-                       PJ5q(env().getGrid());
+        envGetTmp(LatticeComplex, axial_defect);
+        envGetTmp(LatticeComplex, PJ5q);
         std::vector<TComplex> axial_buf;
 
         // Compute <P|D_mu A_mu>, D is backwards derivative.

extras/Hadrons/Modules/MContraction/WeakHamiltonian.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonian.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -100,11 +101,13 @@ public:\
     /* dependency relation */ \
     virtual std::vector<std::string> getInput(void);\
     virtual std::vector<std::string> getOutput(void);\
+public:\
+    std::vector<std::string> VA_label = {"V", "A"};\
+protected:\
     /* setup */ \
     virtual void setup(void);\
     /* execution */ \
     virtual void execute(void);\
-    std::vector<std::string> VA_label = {"V", "A"};\
 };\
 MODULE_REGISTER_NS(modname, T##modname, MContraction);
 

extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.cc

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.cc
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -76,7 +77,7 @@ std::vector<std::string> TWeakHamiltonianEye::getInput(void)
 
 std::vector<std::string> TWeakHamiltonianEye::getOutput(void)
 {
-    std::vector<std::string> out = {getName()};
+    std::vector<std::string> out = {};
     
     return out;
 }
@@ -84,7 +85,15 @@ std::vector<std::string> TWeakHamiltonianEye::getOutput(void)
 // setup ///////////////////////////////////////////////////////////////////////
 void TWeakHamiltonianEye::setup(void)
 {
+    unsigned int ndim = env().getNd();
 
+    envTmpLat(LatticeComplex,  "expbuf");
+    envTmpLat(PropagatorField, "tmp1");
+    envTmpLat(LatticeComplex,  "tmp2");
+    envTmp(std::vector<PropagatorField>, "S_body", 1, ndim, PropagatorField(env().getGrid()));
+    envTmp(std::vector<PropagatorField>, "S_loop", 1, ndim, PropagatorField(env().getGrid()));
+    envTmp(std::vector<LatticeComplex>,  "E_body", 1, ndim, LatticeComplex(env().getGrid()));
+    envTmp(std::vector<LatticeComplex>,  "E_loop", 1, ndim, LatticeComplex(env().getGrid()));
 }
 
 // execution ///////////////////////////////////////////////////////////////////
@@ -95,23 +104,23 @@ void TWeakHamiltonianEye::execute(void)
                  << par().q2 << ", '" << par().q3 << "' and '" << par().q4 
                  << "'." << std::endl;
 
-    CorrWriter             writer(par().output);
+    ResultWriter           writer(RESULT_FILE_NAME(par().output));
-    SlicedPropagator &q1 = *env().template getObject<SlicedPropagator>(par().q1);
+    auto                   &q1 = envGet(SlicedPropagator, par().q1);
-    PropagatorField  &q2 = *env().template getObject<PropagatorField>(par().q2);
+    auto                   &q2 = envGet(PropagatorField, par().q2);
-    PropagatorField  &q3 = *env().template getObject<PropagatorField>(par().q3);
+    auto                   &q3 = envGet(PropagatorField, par().q3);
-    PropagatorField  &q4 = *env().template getObject<PropagatorField>(par().q4);
+    auto                   &q4 = envGet(PropagatorField, par().q4);
-    Gamma g5             = Gamma(Gamma::Algebra::Gamma5);
+    Gamma                  g5  = Gamma(Gamma::Algebra::Gamma5);
-    LatticeComplex         expbuf(env().getGrid());
     std::vector<TComplex>  corrbuf;
     std::vector<Result>    result(n_eye_diag);
     unsigned int ndim    = env().getNd();
 
-    PropagatorField              tmp1(env().getGrid());
+    envGetTmp(LatticeComplex,               expbuf); 
-    LatticeComplex               tmp2(env().getGrid());
+    envGetTmp(PropagatorField,              tmp1);
-    std::vector<PropagatorField> S_body(ndim, tmp1);
+    envGetTmp(LatticeComplex,               tmp2);
-    std::vector<PropagatorField> S_loop(ndim, tmp1);
+    envGetTmp(std::vector<PropagatorField>, S_body);
-    std::vector<LatticeComplex>  E_body(ndim, tmp2);
+    envGetTmp(std::vector<PropagatorField>, S_loop);
-    std::vector<LatticeComplex>  E_loop(ndim, tmp2);
+    envGetTmp(std::vector<LatticeComplex>,  E_body);
+    envGetTmp(std::vector<LatticeComplex>,  E_loop);
 
     // Get sink timeslice of q1.
     SitePropagator q1Snk = q1[par().tSnk];

extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianEye.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by

extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.cc

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.cc
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -76,7 +77,7 @@ std::vector<std::string> TWeakHamiltonianNonEye::getInput(void)
 
 std::vector<std::string> TWeakHamiltonianNonEye::getOutput(void)
 {
-    std::vector<std::string> out = {getName()};
+    std::vector<std::string> out = {};
     
     return out;
 }
@@ -84,7 +85,15 @@ std::vector<std::string> TWeakHamiltonianNonEye::getOutput(void)
 // setup ///////////////////////////////////////////////////////////////////////
 void TWeakHamiltonianNonEye::setup(void)
 {
+    unsigned int ndim = env().getNd();
 
+    envTmpLat(LatticeComplex,  "expbuf");
+    envTmpLat(PropagatorField, "tmp1");
+    envTmpLat(LatticeComplex,  "tmp2");
+    envTmp(std::vector<PropagatorField>, "C_i_side_loop", 1, ndim, PropagatorField(env().getGrid()));
+    envTmp(std::vector<PropagatorField>, "C_f_side_loop", 1, ndim, PropagatorField(env().getGrid()));
+    envTmp(std::vector<LatticeComplex>,  "W_i_side_loop", 1, ndim, LatticeComplex(env().getGrid()));
+    envTmp(std::vector<LatticeComplex>,  "W_f_side_loop", 1, ndim, LatticeComplex(env().getGrid()));
 }
 
 // execution ///////////////////////////////////////////////////////////////////
@@ -95,23 +104,23 @@ void TWeakHamiltonianNonEye::execute(void)
                  << par().q2 << ", '" << par().q3 << "' and '" << par().q4 
                  << "'." << std::endl;
     
-    CorrWriter             writer(par().output);
+    ResultWriter          writer(RESULT_FILE_NAME(par().output));
-    PropagatorField &q1 = *env().template getObject<PropagatorField>(par().q1);
+    auto                  &q1 = envGet(PropagatorField, par().q1);
-    PropagatorField &q2 = *env().template getObject<PropagatorField>(par().q2);
+    auto                  &q2 = envGet(PropagatorField, par().q2);
-    PropagatorField &q3 = *env().template getObject<PropagatorField>(par().q3);
+    auto                  &q3 = envGet(PropagatorField, par().q3);
-    PropagatorField &q4 = *env().template getObject<PropagatorField>(par().q4);
+    auto                  &q4 = envGet(PropagatorField, par().q4);
-    Gamma g5            = Gamma(Gamma::Algebra::Gamma5);
+    Gamma                 g5  = Gamma(Gamma::Algebra::Gamma5);
-    LatticeComplex        expbuf(env().getGrid());
     std::vector<TComplex> corrbuf;
     std::vector<Result>   result(n_noneye_diag); 
-    unsigned int ndim   = env().getNd();
+    unsigned int          ndim = env().getNd();
 
-    PropagatorField              tmp1(env().getGrid());
+    envGetTmp(LatticeComplex,               expbuf); 
-    LatticeComplex               tmp2(env().getGrid());
+    envGetTmp(PropagatorField,              tmp1);
-    std::vector<PropagatorField> C_i_side_loop(ndim, tmp1);
+    envGetTmp(LatticeComplex,               tmp2);
-    std::vector<PropagatorField> C_f_side_loop(ndim, tmp1);
+    envGetTmp(std::vector<PropagatorField>, C_i_side_loop);
-    std::vector<LatticeComplex>  W_i_side_loop(ndim, tmp2);
+    envGetTmp(std::vector<PropagatorField>, C_f_side_loop);
-    std::vector<LatticeComplex>  W_f_side_loop(ndim, tmp2);
+    envGetTmp(std::vector<LatticeComplex>,  W_i_side_loop);
+    envGetTmp(std::vector<LatticeComplex>,  W_f_side_loop);
 
     // Setup for C-type contractions.
     for (int mu = 0; mu < ndim; ++mu)

extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/WeakHamiltonianNonEye.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by

extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.cc

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.cc
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -78,7 +79,7 @@ std::vector<std::string> TWeakNeutral4ptDisc::getInput(void)
 
 std::vector<std::string> TWeakNeutral4ptDisc::getOutput(void)
 {
-    std::vector<std::string> out = {getName()};
+    std::vector<std::string> out = {};
     
     return out;
 }
@@ -86,7 +87,13 @@ std::vector<std::string> TWeakNeutral4ptDisc::getOutput(void)
 // setup ///////////////////////////////////////////////////////////////////////
 void TWeakNeutral4ptDisc::setup(void)
 {
+    unsigned int ndim = env().getNd();
 
+    envTmpLat(LatticeComplex,  "expbuf");
+    envTmpLat(PropagatorField, "tmp");
+    envTmpLat(LatticeComplex,  "curr");
+    envTmp(std::vector<PropagatorField>, "meson", 1, ndim, PropagatorField(env().getGrid()));
+    envTmp(std::vector<PropagatorField>, "loop", 1, ndim,  PropagatorField(env().getGrid()));
 }
 
 // execution ///////////////////////////////////////////////////////////////////
@@ -97,21 +104,21 @@ void TWeakNeutral4ptDisc::execute(void)
                  << par().q2 << ", '" << par().q3 << "' and '" << par().q4 
                  << "'." << std::endl;
 
-    CorrWriter             writer(par().output);
+    ResultWriter          writer(RESULT_FILE_NAME(par().output));
-    PropagatorField &q1 = *env().template getObject<PropagatorField>(par().q1);
+    auto                  &q1 = envGet(PropagatorField, par().q1);
-    PropagatorField &q2 = *env().template getObject<PropagatorField>(par().q2);
+    auto                  &q2 = envGet(PropagatorField, par().q2);
-    PropagatorField &q3 = *env().template getObject<PropagatorField>(par().q3);
+    auto                  &q3 = envGet(PropagatorField, par().q3);
-    PropagatorField &q4 = *env().template getObject<PropagatorField>(par().q4);
+    auto                  &q4 = envGet(PropagatorField, par().q4);
-    Gamma g5            = Gamma(Gamma::Algebra::Gamma5);
+    Gamma                 g5  = Gamma(Gamma::Algebra::Gamma5);
-    LatticeComplex        expbuf(env().getGrid());
     std::vector<TComplex> corrbuf;
     std::vector<Result>   result(n_neut_disc_diag);
-    unsigned int ndim   = env().getNd();
+    unsigned int          ndim = env().getNd();
 
-    PropagatorField              tmp(env().getGrid());
+    envGetTmp(LatticeComplex,               expbuf); 
-    std::vector<PropagatorField> meson(ndim, tmp);
+    envGetTmp(PropagatorField,              tmp);
-    std::vector<PropagatorField> loop(ndim, tmp);
+    envGetTmp(LatticeComplex,               curr);
-    LatticeComplex               curr(env().getGrid());
+    envGetTmp(std::vector<PropagatorField>, meson);
+    envGetTmp(std::vector<PropagatorField>, loop);
 
     // Setup for type 1 contractions.
     for (int mu = 0; mu < ndim; ++mu)

extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MContraction/WeakNeutral4ptDisc.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by

extras/Hadrons/Modules/MFermion/GaugeProp.hpp

@@ -4,12 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MFermion/GaugeProp.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
-Copyright (C) 2017
 
 Author: Antonin Portelli <antonin.portelli@me.com>
-        Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -85,6 +83,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -127,10 +126,13 @@ template <typename FImpl>
 void TGaugeProp<FImpl>::setup(void)
 {
     Ls_ = env().getObjectLs(par().solver);
-    env().template registerLattice<PropagatorField>(getName());
+    envCreateLat(PropagatorField, getName());
+    envTmpLat(FermionField, "source", Ls_);
+    envTmpLat(FermionField, "sol", Ls_);
+    envTmpLat(FermionField, "tmp");
     if (Ls_ > 1)
     {
-        env().template registerLattice<PropagatorField>(getName() + "_5d", Ls_);
+        envCreateLat(PropagatorField, getName() + "_5d", Ls_);
     }
 }
 
@@ -139,26 +141,23 @@ template <typename FImpl>
 void TGaugeProp<FImpl>::execute(void)
 {
     LOG(Message) << "Computing quark propagator '" << getName() << "'"
-    << std::endl;
+                 << std::endl;
     
-    FermionField    source(env().getGrid(Ls_)), sol(env().getGrid(Ls_)),
+    std::string propName = (Ls_ == 1) ? getName() : (getName() + "_5d");
-    tmp(env().getGrid());
+    auto        &prop    = envGet(PropagatorField, propName);
-    std::string     propName = (Ls_ == 1) ? getName() : (getName() + "_5d");
+    auto        &fullSrc = envGet(PropagatorField, par().source);
-    PropagatorField &prop    = *env().template createLattice<PropagatorField>(propName);
+    auto        &solver  = envGet(SolverFn, par().solver);
-    PropagatorField &fullSrc = *env().template getObject<PropagatorField>(par().source);
-    SolverFn        &solver  = *env().template getObject<SolverFn>(par().solver);
-    if (Ls_ > 1)
-    {
-        env().template createLattice<PropagatorField>(getName());
-    }
     
+    envGetTmp(FermionField, source);
+    envGetTmp(FermionField, sol);
+    envGetTmp(FermionField, tmp);
     LOG(Message) << "Inverting using solver '" << par().solver
-    << "' on source '" << par().source << "'" << std::endl;
+                 << "' on source '" << par().source << "'" << std::endl;
     for (unsigned int s = 0; s < Ns; ++s)
     for (unsigned int c = 0; c < Nc; ++c)
     {
         LOG(Message) << "Inversion for spin= " << s << ", color= " << c
-        << std::endl;
+                     << std::endl;
         // source conversion for 4D sources
         if (!env().isObject5d(par().source))
         {
@@ -177,7 +176,7 @@ void TGaugeProp<FImpl>::execute(void)
         {
             if (Ls_ != env().getObjectLs(par().source))
             {
-                HADRON_ERROR("Ls mismatch between quark action and source");
+                HADRON_ERROR(Size, "Ls mismatch between quark action and source");
             }
             else
             {
@@ -190,8 +189,7 @@ void TGaugeProp<FImpl>::execute(void)
         // create 4D propagators from 5D one if necessary
         if (Ls_ > 1)
         {
-            PropagatorField &p4d =
+            PropagatorField &p4d = envGet(PropagatorField, getName());
-                *env().template getObject<PropagatorField>(getName());
             make_4D(sol, tmp, Ls_);
             FermToProp(p4d, tmp, s, c);
         }

extras/Hadrons/Modules/MGauge/Random.cc

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MGauge/Random.cc
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -44,7 +43,9 @@ TRandom::TRandom(const std::string name)
 // dependencies/products ///////////////////////////////////////////////////////
 std::vector<std::string> TRandom::getInput(void)
 {
-    return std::vector<std::string>();
+    std::vector<std::string> in;
+    
+    return in;
 }
 
 std::vector<std::string> TRandom::getOutput(void)
@@ -57,13 +58,14 @@ std::vector<std::string> TRandom::getOutput(void)
 // setup ///////////////////////////////////////////////////////////////////////
 void TRandom::setup(void)
 {
-    env().registerLattice<LatticeGaugeField>(getName());
+    envCreateLat(LatticeGaugeField, getName());
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 void TRandom::execute(void)
 {
     LOG(Message) << "Generating random gauge configuration" << std::endl;
-    LatticeGaugeField &U = *env().createLattice<LatticeGaugeField>(getName());
+    
+    auto &U = envGet(LatticeGaugeField, getName());
     SU3::HotConfiguration(*env().get4dRng(), U);
 }

extras/Hadrons/Modules/MGauge/Random.hpp

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MGauge/Random.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -51,6 +50,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution

extras/Hadrons/Modules/MGauge/StochEm.cc

@@ -4,9 +4,9 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MGauge/StochEm.cc
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
+Author: Antonin Portelli <antonin.portelli@me.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -57,32 +57,28 @@ std::vector<std::string> TStochEm::getOutput(void)
 // setup ///////////////////////////////////////////////////////////////////////
 void TStochEm::setup(void)
 {
-    if (!env().hasRegisteredObject("_" + getName() + "_weight"))
+    if (!env().hasCreatedObject("_" + getName() + "_weight"))
     {
-        env().registerLattice<EmComp>("_" + getName() + "_weight");
+        envCacheLat(EmComp, "_" + getName() + "_weight");
     }
-    env().registerLattice<EmField>(getName());
+    envCreateLat(EmField, getName());
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 void TStochEm::execute(void)
 {
+    LOG(Message) << "Generating stochatic EM potential..." << std::endl;
+
     PhotonR photon(par().gauge, par().zmScheme);
-    EmField &a = *env().createLattice<EmField>(getName());
+    auto    &a = envGet(EmField, getName());
-    EmComp  *w;
+    auto    &w = envGet(EmComp, "_" + getName() + "_weight");
     
     if (!env().hasCreatedObject("_" + getName() + "_weight"))
     {
         LOG(Message) << "Caching stochatic EM potential weight (gauge: "
                      << par().gauge << ", zero-mode scheme: "
                      << par().zmScheme << ")..." << std::endl;
-        w = env().createLattice<EmComp>("_" + getName() + "_weight");
+        photon.StochasticWeight(w);
-        photon.StochasticWeight(*w);
     }
-    else
+    photon.StochasticField(a, *env().get4dRng(), w);
-    {
-        w = env().getObject<EmComp>("_" + getName() + "_weight");
-    }
-    LOG(Message) << "Generating stochatic EM potential..." << std::endl;
-    photon.StochasticField(a, *env().get4dRng(), *w);
 }

extras/Hadrons/Modules/MGauge/StochEm.hpp

@@ -4,9 +4,9 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MGauge/StochEm.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
+Author: Antonin Portelli <antonin.portelli@me.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -60,6 +60,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution

extras/Hadrons/Modules/MGauge/Unit.cc

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MGauge/Unit.cc
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -57,13 +56,14 @@ std::vector<std::string> TUnit::getOutput(void)
 // setup ///////////////////////////////////////////////////////////////////////
 void TUnit::setup(void)
 {
-    env().registerLattice<LatticeGaugeField>(getName());
+    envCreateLat(LatticeGaugeField, getName());
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 void TUnit::execute(void)
 {
     LOG(Message) << "Creating unit gauge configuration" << std::endl;
-    LatticeGaugeField &U = *env().createLattice<LatticeGaugeField>(getName());
+    
+    auto &U = envGet(LatticeGaugeField, getName());
     SU3::ColdConfiguration(*env().get4dRng(), U);
 }

extras/Hadrons/Modules/MGauge/Unit.hpp

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MGauge/Unit.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -51,6 +50,7 @@ public:
     // dependencies/products
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution

extras/Hadrons/Modules/MIO/LoadBinary.hpp

@@ -0,0 +1,140 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MIO/LoadBinary.hpp
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef Hadrons_MIO_LoadBinary_hpp_
+#define Hadrons_MIO_LoadBinary_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                       Load a binary configurations                         *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MIO)
+
+class LoadBinaryPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(LoadBinaryPar,
+                                    std::string, file,
+                                    std::string, format);
+};
+
+template <typename Impl>
+class TLoadBinary: public Module<LoadBinaryPar>
+{
+public:
+    typedef typename Impl::Field                  Field;
+    typedef typename Impl::Simd                   Simd;
+    typedef typename Field::vector_object         vobj;
+    typedef typename vobj::scalar_object          sobj;
+    typedef typename sobj::DoublePrecision        sobj_double;
+    typedef BinarySimpleMunger<sobj_double, sobj> Munger;
+public:
+    // constructor
+    TLoadBinary(const std::string name);
+    // destructor
+    virtual ~TLoadBinary(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+};
+
+MODULE_REGISTER_NS(LoadBinary, TLoadBinary<GIMPL>, MIO);
+MODULE_REGISTER_NS(LoadBinaryScalarSU2, TLoadBinary<ScalarNxNAdjImplR<2>>, MIO);
+MODULE_REGISTER_NS(LoadBinaryScalarSU3, TLoadBinary<ScalarNxNAdjImplR<3>>, MIO);
+MODULE_REGISTER_NS(LoadBinaryScalarSU4, TLoadBinary<ScalarNxNAdjImplR<4>>, MIO);
+MODULE_REGISTER_NS(LoadBinaryScalarSU5, TLoadBinary<ScalarNxNAdjImplR<5>>, MIO);
+MODULE_REGISTER_NS(LoadBinaryScalarSU6, TLoadBinary<ScalarNxNAdjImplR<6>>, MIO);
+
+/******************************************************************************
+ *                         TLoadBinary implementation                         *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename Impl>
+TLoadBinary<Impl>::TLoadBinary(const std::string name)
+: Module<LoadBinaryPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename Impl>
+std::vector<std::string> TLoadBinary<Impl>::getInput(void)
+{
+    std::vector<std::string> in;
+    
+    return in;
+}
+
+template <typename Impl>
+std::vector<std::string> TLoadBinary<Impl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename Impl>
+void TLoadBinary<Impl>::setup(void)
+{
+    envCreateLat(Field, getName());
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename Impl>
+void TLoadBinary<Impl>::execute(void)
+{
+    Munger      munge;
+    uint32_t    nersc_csum, scidac_csuma, scidac_csumb;
+    auto        &U = envGet(Field, getName());
+    std::string filename = par().file + "."
+                           + std::to_string(vm().getTrajectory());
+
+    LOG(Message) << "Loading " << par().format 
+                 << " binary configuration from file '" << filename
+                 << "'" << std::endl;
+    BinaryIO::readLatticeObject<vobj, sobj_double>(U, filename, munge, 0, 
+                                                   par().format, nersc_csum,
+                                                   scidac_csuma, scidac_csumb);
+    LOG(Message) << "Checksums:" << std::endl;
+    LOG(Message) << "  NERSC    " << nersc_csum << std::endl;
+    LOG(Message) << "  SciDAC A " << scidac_csuma << std::endl;
+    LOG(Message) << "  SciDAC B " << scidac_csumb << std::endl;
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MIO_LoadBinary_hpp_

extras/Hadrons/Modules/MIO/LoadNersc.cc

@@ -2,10 +2,9 @@
 
 Grid physics library, www.github.com/paboyle/Grid 
 
-Source file: extras/Hadrons/Modules/MGauge/Load.cc
+Source file: extras/Hadrons/Modules/MIO/LoadNersc.cc
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -26,30 +25,29 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-
+#include <Grid/Hadrons/Modules/MIO/LoadNersc.hpp>
-#include <Grid/Hadrons/Modules/MGauge/Load.hpp>
 
 using namespace Grid;
 using namespace Hadrons;
-using namespace MGauge;
+using namespace MIO;
 
 /******************************************************************************
-*                           TLoad implementation                               *
+*                       TLoadNersc implementation                             *
 ******************************************************************************/
 // constructor /////////////////////////////////////////////////////////////////
-TLoad::TLoad(const std::string name)
+TLoadNersc::TLoadNersc(const std::string name)
-: Module<LoadPar>(name)
+: Module<LoadNerscPar>(name)
 {}
 
 // dependencies/products ///////////////////////////////////////////////////////
-std::vector<std::string> TLoad::getInput(void)
+std::vector<std::string> TLoadNersc::getInput(void)
 {
     std::vector<std::string> in;
     
     return in;
 }
 
-std::vector<std::string> TLoad::getOutput(void)
+std::vector<std::string> TLoadNersc::getOutput(void)
 {
     std::vector<std::string> out = {getName()};
     
@@ -57,21 +55,21 @@ std::vector<std::string> TLoad::getOutput(void)
 }
 
 // setup ///////////////////////////////////////////////////////////////////////
-void TLoad::setup(void)
+void TLoadNersc::setup(void)
 {
-    env().registerLattice<LatticeGaugeField>(getName());
+    envCreateLat(LatticeGaugeField, getName());
 }
 
 // execution ///////////////////////////////////////////////////////////////////
-void TLoad::execute(void)
+void TLoadNersc::execute(void)
 {
-    FieldMetaData  header;
+    FieldMetaData header;
-    std::string fileName = par().file + "."
+    std::string   fileName = par().file + "."
-                           + std::to_string(env().getTrajectory());
+                             + std::to_string(vm().getTrajectory());
-    
     LOG(Message) << "Loading NERSC configuration from file '" << fileName
                  << "'" << std::endl;
-    LatticeGaugeField &U = *env().createLattice<LatticeGaugeField>(getName());
+
+    auto &U = envGet(LatticeGaugeField, getName());
     NerscIO::readConfiguration(U, header, fileName);
     LOG(Message) << "NERSC header:" << std::endl;
     dump_meta_data(header, LOG(Message));

extras/Hadrons/Modules/MIO/LoadNersc.hpp

@@ -2,10 +2,9 @@
 
 Grid physics library, www.github.com/paboyle/Grid 
 
-Source file: extras/Hadrons/Modules/MGauge/Load.hpp
+Source file: extras/Hadrons/Modules/MIO/LoadNersc.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -26,9 +25,8 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 See the full license in the file "LICENSE" in the top level distribution directory
 *************************************************************************************/
 /*  END LEGAL */
-
+#ifndef Hadrons_MIO_LoadNersc_hpp_
-#ifndef Hadrons_MGauge_Load_hpp_
+#define Hadrons_MIO_LoadNersc_hpp_
-#define Hadrons_MGauge_Load_hpp_
 
 #include <Grid/Hadrons/Global.hpp>
 #include <Grid/Hadrons/Module.hpp>
@@ -37,24 +35,24 @@ See the full license in the file "LICENSE" in the top level distribution directo
 BEGIN_HADRONS_NAMESPACE
 
 /******************************************************************************
- *                         Load a NERSC configuration                         *
+ *                       Load a NERSC configuration                           *
  ******************************************************************************/
-BEGIN_MODULE_NAMESPACE(MGauge)
+BEGIN_MODULE_NAMESPACE(MIO)
 
-class LoadPar: Serializable
+class LoadNerscPar: Serializable
 {
 public:
-    GRID_SERIALIZABLE_CLASS_MEMBERS(LoadPar,
+    GRID_SERIALIZABLE_CLASS_MEMBERS(LoadNerscPar,
                                     std::string, file);
 };
 
-class TLoad: public Module<LoadPar>
+class TLoadNersc: public Module<LoadNerscPar>
 {
 public:
     // constructor
-    TLoad(const std::string name);
+    TLoadNersc(const std::string name);
     // destructor
-    virtual ~TLoad(void) = default;
+    virtual ~TLoadNersc(void) = default;
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
@@ -64,10 +62,10 @@ public:
     virtual void execute(void);
 };
 
-MODULE_REGISTER_NS(Load, TLoad, MGauge);
+MODULE_REGISTER_NS(LoadNersc, TLoadNersc, MIO);
 
 END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE
 
-#endif // Hadrons_MGauge_Load_hpp_
+#endif // Hadrons_MIO_LoadNersc_hpp_

extras/Hadrons/Modules/MLoop/NoiseLoop.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MLoop/NoiseLoop.hpp
 
-Copyright (C) 2016
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -74,6 +75,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -112,16 +114,16 @@ std::vector<std::string> TNoiseLoop<FImpl>::getOutput(void)
 template <typename FImpl>
 void TNoiseLoop<FImpl>::setup(void)
 {
-    env().template registerLattice<PropagatorField>(getName());
+    envCreateLat(PropagatorField, getName());
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 template <typename FImpl>
 void TNoiseLoop<FImpl>::execute(void)
 {
-    PropagatorField &loop = *env().template createLattice<PropagatorField>(getName());
+    auto &loop = envGet(PropagatorField, getName());
-    PropagatorField &q    = *env().template getObject<PropagatorField>(par().q);
+    auto &q    = envGet(PropagatorField, par().q);
-    PropagatorField &eta  = *env().template getObject<PropagatorField>(par().eta);
+    auto &eta  = envGet(PropagatorField, par().eta);
     loop = q*adj(eta);
 }
 

extras/Hadrons/Modules/MScalar/ChargedProp.cc

@@ -1,3 +1,31 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MScalar/ChargedProp.cc
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: James Harrison <jch1g10@soton.ac.uk>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
 #include <Grid/Hadrons/Modules/MScalar/ChargedProp.hpp>
 #include <Grid/Hadrons/Modules/MScalar/Scalar.hpp>
 
@@ -37,90 +65,44 @@ void TChargedProp::setup(void)
     {
         phaseName_.push_back("_shiftphase_" + std::to_string(mu));
     }
-    GFSrcName_ = "_" + getName() + "_DinvSrc";
+    GFSrcName_ = getName() + "_DinvSrc";
-    if (!env().hasRegisteredObject(freeMomPropName_))
+    fftName_   = getName() + "_fft";
+
+    freeMomPropDone_ = env().hasCreatedObject(freeMomPropName_);
+    GFSrcDone_       = env().hasCreatedObject(GFSrcName_);
+    phasesDone_      = env().hasCreatedObject(phaseName_[0]);
+    envCacheLat(ScalarField, freeMomPropName_);
+    for (unsigned int mu = 0; mu < env().getNd(); ++mu)
     {
-        env().registerLattice<ScalarField>(freeMomPropName_);
+        envCacheLat(ScalarField, phaseName_[mu]);
     }
-    if (!env().hasRegisteredObject(phaseName_[0]))
+    envCacheLat(ScalarField, GFSrcName_);
-    {
+    envCreateLat(ScalarField, getName());
-        for (unsigned int mu = 0; mu < env().getNd(); ++mu)
+    envTmpLat(ScalarField, "buf");
-        {
+    envTmpLat(ScalarField, "result");
-            env().registerLattice<ScalarField>(phaseName_[mu]);
+    envTmpLat(ScalarField, "Amu");
-        }
+    envCache(FFT, fftName_, 1, env().getGrid());
-    }
-    if (!env().hasRegisteredObject(GFSrcName_))
-    {
-        env().registerLattice<ScalarField>(GFSrcName_);
-    }
-    env().registerLattice<ScalarField>(getName());
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 void TChargedProp::execute(void)
 {
     // CACHING ANALYTIC EXPRESSIONS
-    ScalarField &source = *env().getObject<ScalarField>(par().source);
+    makeCaches();
-    Complex     ci(0.0,1.0);
-    FFT         fft(env().getGrid());
-    
-    // cache free scalar propagator
-    if (!env().hasCreatedObject(freeMomPropName_))
-    {
-        LOG(Message) << "Caching momentum space free scalar propagator"
-                     << " (mass= " << par().mass << ")..." << std::endl;
-        freeMomProp_ = env().createLattice<ScalarField>(freeMomPropName_);
-        SIMPL::MomentumSpacePropagator(*freeMomProp_, par().mass);
-    }
-    else
-    {
-        freeMomProp_ = env().getObject<ScalarField>(freeMomPropName_);
-    }
-    // cache G*F*src
-    if (!env().hasCreatedObject(GFSrcName_))
-        
-    {
-        GFSrc_ = env().createLattice<ScalarField>(GFSrcName_);
-        fft.FFT_all_dim(*GFSrc_, source, FFT::forward);
-        *GFSrc_ = (*freeMomProp_)*(*GFSrc_);
-    }
-    else
-    {
-        GFSrc_ = env().getObject<ScalarField>(GFSrcName_);
-    }
-    // cache phases
-    if (!env().hasCreatedObject(phaseName_[0]))
-    {
-        std::vector<int> &l = env().getGrid()->_fdimensions;
-        
-        LOG(Message) << "Caching shift phases..." << std::endl;
-        for (unsigned int mu = 0; mu < env().getNd(); ++mu)
-        {
-            Real    twoPiL = M_PI*2./l[mu];
-            
-            phase_.push_back(env().createLattice<ScalarField>(phaseName_[mu]));
-            LatticeCoordinate(*(phase_[mu]), mu);
-            *(phase_[mu]) = exp(ci*twoPiL*(*(phase_[mu])));
-        }
-    }
-    else
-    {
-        for (unsigned int mu = 0; mu < env().getNd(); ++mu)
-        {
-            phase_.push_back(env().getObject<ScalarField>(phaseName_[mu]));
-        }
-    }
 
     // PROPAGATOR CALCULATION
     LOG(Message) << "Computing charged scalar propagator"
                  << " (mass= " << par().mass
                  << ", charge= " << par().charge << ")..." << std::endl;
     
-    ScalarField &prop   = *env().createLattice<ScalarField>(getName());
+    auto   &prop  = envGet(ScalarField, getName());
-    ScalarField buf(env().getGrid());
+    auto   &GFSrc = envGet(ScalarField, GFSrcName_);
-    ScalarField &GFSrc = *GFSrc_, &G = *freeMomProp_;
+    auto   &G     = envGet(ScalarField, freeMomPropName_);
-    double      q = par().charge;
+    auto   &fft   = envGet(FFT, fftName_);
-    
+    double q      = par().charge;
+    envGetTmp(ScalarField, result); 
+    envGetTmp(ScalarField, buf); 
+
     // G*F*Src
     prop = GFSrc;
 
@@ -146,12 +128,12 @@ void TChargedProp::execute(void)
     if (!par().output.empty())
     {
         std::string           filename = par().output + "." +
-                                         std::to_string(env().getTrajectory());
+                                         std::to_string(vm().getTrajectory());
         
         LOG(Message) << "Saving zero-momentum projection to '"
                      << filename << "'..." << std::endl;
         
-        CorrWriter            writer(filename);
+        ResultWriter          writer(RESULT_FILE_NAME(par().output));
         std::vector<TComplex> vecBuf;
         std::vector<Complex>  result;
         
@@ -166,15 +148,55 @@ void TChargedProp::execute(void)
     }
 }
 
+void TChargedProp::makeCaches(void)
+{
+    auto &freeMomProp = envGet(ScalarField, freeMomPropName_);
+    auto &GFSrc       = envGet(ScalarField, GFSrcName_);
+    auto &fft         = envGet(FFT, fftName_);
+
+    if (!freeMomPropDone_)
+    {
+        LOG(Message) << "Caching momentum space free scalar propagator"
+                     << " (mass= " << par().mass << ")..." << std::endl;
+        SIMPL::MomentumSpacePropagator(freeMomProp, par().mass);
+    }
+    if (!GFSrcDone_)
+    {   
+        FFT  fft(env().getGrid());
+        auto &source = envGet(ScalarField, par().source);
+
+        LOG(Message) << "Caching G*F*src..." << std::endl;
+        fft.FFT_all_dim(GFSrc, source, FFT::forward);
+        GFSrc = freeMomProp*GFSrc;
+    }
+    if (!phasesDone_)
+    {
+        std::vector<int> &l = env().getGrid()->_fdimensions;
+        Complex          ci(0.0,1.0);
+        
+        LOG(Message) << "Caching shift phases..." << std::endl;
+        for (unsigned int mu = 0; mu < env().getNd(); ++mu)
+        {
+            Real twoPiL = M_PI*2./l[mu];
+            auto &phmu  = envGet(ScalarField, phaseName_[mu]);
+            
+            LatticeCoordinate(phmu, mu);
+            phmu = exp(ci*twoPiL*phmu);
+            phase_.push_back(&phmu);
+        }
+    }
+}
+
 void TChargedProp::momD1(ScalarField &s, FFT &fft)
 {
-    EmField     &A = *env().getObject<EmField>(par().emField);
+    auto        &A = envGet(EmField, par().emField);
-    ScalarField buf(env().getGrid()), result(env().getGrid()),
-                Amu(env().getGrid());
     Complex     ci(0.0,1.0);
 
-    result = zero;
+    envGetTmp(ScalarField, buf);
+    envGetTmp(ScalarField, result);
+    envGetTmp(ScalarField, Amu);
 
+    result = zero;
     for (unsigned int mu = 0; mu < env().getNd(); ++mu)
     {
         Amu = peekLorentz(A, mu);
@@ -198,12 +220,13 @@ void TChargedProp::momD1(ScalarField &s, FFT &fft)
 
 void TChargedProp::momD2(ScalarField &s, FFT &fft)
 {
-    EmField     &A = *env().getObject<EmField>(par().emField);
+    auto &A = envGet(EmField, par().emField);
-    ScalarField buf(env().getGrid()), result(env().getGrid()),
+
-                Amu(env().getGrid());
+    envGetTmp(ScalarField, buf);
+    envGetTmp(ScalarField, result);
+    envGetTmp(ScalarField, Amu);
 
     result = zero;
-    
     for (unsigned int mu = 0; mu < env().getNd(); ++mu)
     {
         Amu = peekLorentz(A, mu);

extras/Hadrons/Modules/MScalar/ChargedProp.hpp

@@ -1,3 +1,30 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MScalar/ChargedProp.hpp
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
 #ifndef Hadrons_MScalar_ChargedProp_hpp_
 #define Hadrons_MScalar_ChargedProp_hpp_
 
@@ -37,19 +64,20 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
     virtual void execute(void);
 private:
+    void makeCaches(void);
     void momD1(ScalarField &s, FFT &fft);
     void momD2(ScalarField &s, FFT &fft);
 private:
-    std::string                freeMomPropName_, GFSrcName_;
+    bool                       freeMomPropDone_, GFSrcDone_, phasesDone_;
+    std::string                freeMomPropName_, GFSrcName_, fftName_;
     std::vector<std::string>   phaseName_;
-    ScalarField                *freeMomProp_, *GFSrc_;
     std::vector<ScalarField *> phase_;
-    EmField                    *A;
 };
 
 MODULE_REGISTER_NS(ChargedProp, TChargedProp, MScalar);

extras/Hadrons/Modules/MScalar/FreeProp.cc

@@ -1,3 +1,30 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MScalar/FreeProp.cc
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
 #include <Grid/Hadrons/Modules/MScalar/FreeProp.hpp>
 #include <Grid/Hadrons/Modules/MScalar/Scalar.hpp>
 
@@ -33,38 +60,31 @@ void TFreeProp::setup(void)
 {
     freeMomPropName_ = FREEMOMPROP(par().mass);
     
-    if (!env().hasRegisteredObject(freeMomPropName_))
+    freePropDone_ = env().hasCreatedObject(freeMomPropName_);
-    {
+    envCacheLat(ScalarField, freeMomPropName_);
-        env().registerLattice<ScalarField>(freeMomPropName_);
+    envCreateLat(ScalarField, getName());
-    }
-    env().registerLattice<ScalarField>(getName());
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 void TFreeProp::execute(void)
 {
-    ScalarField &prop   = *env().createLattice<ScalarField>(getName());
+    auto &freeMomProp = envGet(ScalarField, freeMomPropName_);
-    ScalarField &source = *env().getObject<ScalarField>(par().source);
+    auto &prop        = envGet(ScalarField, getName());
-    ScalarField *freeMomProp;
+    auto &source      = envGet(ScalarField, par().source);
 
-    if (!env().hasCreatedObject(freeMomPropName_))
+    if (!freePropDone_)
     {
         LOG(Message) << "Caching momentum space free scalar propagator"
                      << " (mass= " << par().mass << ")..." << std::endl;
-        freeMomProp = env().createLattice<ScalarField>(freeMomPropName_);
+        SIMPL::MomentumSpacePropagator(freeMomProp, par().mass);
-        SIMPL::MomentumSpacePropagator(*freeMomProp, par().mass);
-    }
-    else
-    {
-        freeMomProp = env().getObject<ScalarField>(freeMomPropName_);
     }
     LOG(Message) << "Computing free scalar propagator..." << std::endl;
-    SIMPL::FreePropagator(source, prop, *freeMomProp);
+    SIMPL::FreePropagator(source, prop, freeMomProp);
     
     if (!par().output.empty())
     {
         TextWriter            writer(par().output + "." +
-                                     std::to_string(env().getTrajectory()));
+                                     std::to_string(vm().getTrajectory()));
         std::vector<TComplex> buf;
         std::vector<Complex>  result;
         

extras/Hadrons/Modules/MScalar/FreeProp.hpp

@@ -1,3 +1,30 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MScalar/FreeProp.hpp
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
 #ifndef Hadrons_MScalar_FreeProp_hpp_
 #define Hadrons_MScalar_FreeProp_hpp_
 
@@ -33,12 +60,14 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
     virtual void execute(void);
 private:
     std::string freeMomPropName_;
+    bool        freePropDone_;
 };
 
 MODULE_REGISTER_NS(FreeProp, TFreeProp, MScalar);

extras/Hadrons/Modules/MScalar/Scalar.hpp

@@ -1,3 +1,30 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MScalar/Scalar.hpp
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
 #ifndef Hadrons_Scalar_hpp_
 #define Hadrons_Scalar_hpp_
 

extras/Hadrons/Modules/MScalarSUN/TrMag.hpp

@@ -0,0 +1,146 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MScalarSUN/TrMag.hpp
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef Hadrons_MScalarSUN_TrMag_hpp_
+#define Hadrons_MScalarSUN_TrMag_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                       Module to compute tr(mag^n)                          *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MScalarSUN)
+
+class TrMagPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(TrMagPar,
+                                    std::string,  field,
+                                    unsigned int, maxPow,
+                                    std::string,  output);
+};
+
+template <typename SImpl>
+class TTrMag: public Module<TrMagPar>
+{
+public:
+    typedef typename SImpl::Field        Field;
+    typedef typename SImpl::ComplexField ComplexField;
+    class Result: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,
+                                        std::string, op,
+                                        Real,        value);
+    };
+public:
+    // constructor
+    TTrMag(const std::string name);
+    // destructor
+    virtual ~TTrMag(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+};
+
+MODULE_REGISTER_NS(TrMagSU2, TTrMag<ScalarNxNAdjImplR<2>>, MScalarSUN);
+MODULE_REGISTER_NS(TrMagSU3, TTrMag<ScalarNxNAdjImplR<3>>, MScalarSUN);
+MODULE_REGISTER_NS(TrMagSU4, TTrMag<ScalarNxNAdjImplR<4>>, MScalarSUN);
+MODULE_REGISTER_NS(TrMagSU5, TTrMag<ScalarNxNAdjImplR<5>>, MScalarSUN);
+MODULE_REGISTER_NS(TrMagSU6, TTrMag<ScalarNxNAdjImplR<6>>, MScalarSUN);
+
+/******************************************************************************
+ *                         TTrMag implementation                              *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename SImpl>
+TTrMag<SImpl>::TTrMag(const std::string name)
+: Module<TrMagPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename SImpl>
+std::vector<std::string> TTrMag<SImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().field};
+    
+    return in;
+}
+
+template <typename SImpl>
+std::vector<std::string> TTrMag<SImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename SImpl>
+void TTrMag<SImpl>::setup(void)
+{}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename SImpl>
+void TTrMag<SImpl>::execute(void)
+{
+    LOG(Message) << "Computing tr(mag^n) for n even up to " << par().maxPow
+                 << "..." << std::endl;
+
+    std::vector<Result> result;
+    ResultWriter        writer(RESULT_FILE_NAME(par().output));
+    auto                &phi = envGet(Field, par().field);
+
+    auto m2 = sum(phi), mn = m2;
+
+    m2 = -m2*m2;
+    mn = 1.;
+    for (unsigned int n = 2; n <= par().maxPow; n += 2)
+    {
+        Result r;
+
+        mn = mn*m2;
+        r.op    = "tr(mag^" + std::to_string(n) + ")";
+        r.value = TensorRemove(trace(mn)).real();
+        result.push_back(r);
+    }
+    write(writer, "trmag", result);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MScalarSUN_TrMag_hpp_

extras/Hadrons/Modules/MScalarSUN/TrPhi.hpp

@@ -0,0 +1,182 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MScalarSUN/TrPhi.hpp
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef Hadrons_MScalarSUN_TrPhi_hpp_
+#define Hadrons_MScalarSUN_TrPhi_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                         Module to compute tr(phi^n)                        *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MScalarSUN)
+
+class TrPhiPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(TrPhiPar,
+                                    std::string,  field,
+                                    unsigned int, maxPow,
+                                    std::string,  output);
+};
+
+template <typename SImpl>
+class TTrPhi: public Module<TrPhiPar>
+{
+public:
+    typedef typename SImpl::Field        Field;
+    typedef typename SImpl::ComplexField ComplexField;
+    class Result: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,
+                                        std::string, op,
+                                        Real,        value);
+    };
+public:
+    // constructor
+    TTrPhi(const std::string name);
+    // destructor
+    virtual ~TTrPhi(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+private:
+    // output name generator
+    std::string outName(const unsigned int n);
+};
+
+MODULE_REGISTER_NS(TrPhiSU2, TTrPhi<ScalarNxNAdjImplR<2>>, MScalarSUN);
+MODULE_REGISTER_NS(TrPhiSU3, TTrPhi<ScalarNxNAdjImplR<3>>, MScalarSUN);
+MODULE_REGISTER_NS(TrPhiSU4, TTrPhi<ScalarNxNAdjImplR<4>>, MScalarSUN);
+MODULE_REGISTER_NS(TrPhiSU5, TTrPhi<ScalarNxNAdjImplR<5>>, MScalarSUN);
+MODULE_REGISTER_NS(TrPhiSU6, TTrPhi<ScalarNxNAdjImplR<6>>, MScalarSUN);
+
+/******************************************************************************
+ *                          TTrPhi implementation                             *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename SImpl>
+TTrPhi<SImpl>::TTrPhi(const std::string name)
+: Module<TrPhiPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename SImpl>
+std::vector<std::string> TTrPhi<SImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().field};
+    
+    return in;
+}
+
+template <typename SImpl>
+std::vector<std::string> TTrPhi<SImpl>::getOutput(void)
+{
+    std::vector<std::string> out;
+
+    for (unsigned int n = 2; n <= par().maxPow; n += 2)
+    {
+        out.push_back(outName(n));
+    }
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename SImpl>
+void TTrPhi<SImpl>::setup(void)
+{
+    if (par().maxPow < 2)
+    {
+        HADRON_ERROR(Size, "'maxPow' should be at least equal to 2");
+    }
+    envTmpLat(Field, "phi2");
+    envTmpLat(Field, "buf");
+    for (unsigned int n = 2; n <= par().maxPow; n += 2)
+    {
+        envCreateLat(ComplexField, outName(n));
+    }
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename SImpl>
+void TTrPhi<SImpl>::execute(void)
+{
+    LOG(Message) << "Computing tr(phi^n) for n even up to " << par().maxPow
+                 << "..." << std::endl; 
+
+    std::vector<Result> result;
+    auto                &phi = envGet(Field, par().field);
+
+    envGetTmp(Field, phi2);
+    envGetTmp(Field, buf);
+    buf  = 1.;
+    phi2 = -phi*phi; 
+    for (unsigned int n = 2; n <= par().maxPow; n += 2)
+    {
+        auto &phin = envGet(ComplexField, outName(n));
+
+        buf  = buf*phi2;
+        phin = trace(buf);
+        if (!par().output.empty())
+        {
+            Result r;
+
+            r.op    = "tr(phi^" + std::to_string(n) + ")";
+            r.value = TensorRemove(sum(phin)).real();
+            result.push_back(r);
+        }
+    }
+    if (result.size() > 0)
+    {
+        ResultWriter writer(RESULT_FILE_NAME(par().output));
+
+        write(writer, "trphi", result);
+    }
+}
+
+// output name generator ///////////////////////////////////////////////////////
+template <typename SImpl>
+std::string TTrPhi<SImpl>::outName(const unsigned int n)
+{
+    return getName() + "_" + std::to_string(n);
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MScalarSUN_TrPhi_hpp_

extras/Hadrons/Modules/MScalarSUN/TwoPoint.hpp

@@ -0,0 +1,184 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/Modules/MScalarSUN/TwoPoint.hpp
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+#ifndef Hadrons_MScalarSUN_TwoPoint_hpp_
+#define Hadrons_MScalarSUN_TwoPoint_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Module.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                 2-pt functions for a given set of operators                *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MScalarSUN)
+
+class TwoPointPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(TwoPointPar,
+                                    std::vector<std::string>, op,
+                                    std::string,              output);
+};
+
+template <typename SImpl>
+class TTwoPoint: public Module<TwoPointPar>
+{
+public:
+    typedef typename SImpl::Field        Field;
+    typedef typename SImpl::ComplexField ComplexField;
+    class Result: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(Result,
+                                        std::string, sink,
+                                        std::string, source,
+                                        std::vector<Complex>, data);
+    };
+public:
+    // constructor
+    TTwoPoint(const std::string name);
+    // destructor
+    virtual ~TTwoPoint(void) = default;
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+private:
+    // make 2-pt function
+    template <class SinkSite, class SourceSite>
+    std::vector<Complex> makeTwoPoint(const std::vector<SinkSite>   &sink,
+                                      const std::vector<SourceSite> &source);
+};
+
+MODULE_REGISTER_NS(TwoPointSU2, TTwoPoint<ScalarNxNAdjImplR<2>>, MScalarSUN);
+MODULE_REGISTER_NS(TwoPointSU3, TTwoPoint<ScalarNxNAdjImplR<3>>, MScalarSUN);
+MODULE_REGISTER_NS(TwoPointSU4, TTwoPoint<ScalarNxNAdjImplR<4>>, MScalarSUN);
+MODULE_REGISTER_NS(TwoPointSU5, TTwoPoint<ScalarNxNAdjImplR<5>>, MScalarSUN);
+MODULE_REGISTER_NS(TwoPointSU6, TTwoPoint<ScalarNxNAdjImplR<6>>, MScalarSUN);
+
+/******************************************************************************
+ *                 TTwoPoint implementation                             *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename SImpl>
+TTwoPoint<SImpl>::TTwoPoint(const std::string name)
+: Module<TwoPointPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename SImpl>
+std::vector<std::string> TTwoPoint<SImpl>::getInput(void)
+{   
+    return par().op;
+}
+
+template <typename SImpl>
+std::vector<std::string> TTwoPoint<SImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {};
+
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename SImpl>
+void TTwoPoint<SImpl>::setup(void)
+{
+    const unsigned int nt = env().getDim().back();
+    envTmp(std::vector<std::vector<TComplex>>, "slicedOp", 1, par().op.size(), 
+           std::vector<TComplex>(nt));
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename SImpl>
+void TTwoPoint<SImpl>::execute(void)
+{
+    LOG(Message) << "Computing 2-point functions for operators:" << std::endl;
+    for (auto &o: par().op)
+    {
+        LOG(Message) << "  '" << o << "'" << std::endl;
+    }
+
+    ResultWriter        writer(RESULT_FILE_NAME(par().output));
+    const unsigned int  nd = env().getDim().size();
+    std::vector<Result> result;
+    
+    envGetTmp(std::vector<std::vector<TComplex>>, slicedOp);
+    for (unsigned int i = 0; i < par().op.size(); ++i)
+    {
+        auto &op = envGet(ComplexField, par().op[i]);
+
+        sliceSum(op, slicedOp[i], nd - 1);
+    }
+    for (unsigned int i = 0; i < par().op.size(); ++i)
+    for (unsigned int j = 0; j < par().op.size(); ++j)
+    {
+        Result r;
+
+        r.sink   = par().op[i];
+        r.source = par().op[j];
+        r.data   = makeTwoPoint(slicedOp[i], slicedOp[j]);
+        result.push_back(r);
+    }
+    write(writer, "twopt", result);
+}
+
+// make 2-pt function //////////////////////////////////////////////////////////
+template <class SImpl>
+template <class SinkSite, class SourceSite>
+std::vector<Complex> TTwoPoint<SImpl>::makeTwoPoint(
+                                  const std::vector<SinkSite>   &sink,
+                                  const std::vector<SourceSite> &source)
+{
+    assert(sink.size() == source.size());
+    
+    unsigned int         nt = sink.size();
+    std::vector<Complex> res(nt, 0.);
+    
+    for (unsigned int dt = 0; dt < nt; ++dt)
+    {
+        for (unsigned int t  = 0; t < nt; ++t)
+        {
+            res[dt] += TensorRemove(trace(sink[(t+dt)%nt]*source[t]));
+        }
+        res[dt] *= 1./static_cast<double>(nt);
+    }
+    
+    return res;
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MScalarSUN_TwoPoint_hpp_

extras/Hadrons/Modules/MSink/Point.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MSink/Point.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
 Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -61,10 +62,14 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
     virtual void execute(void);
+private:
+    bool        hasPhase_{false}; 
+    std::string momphName_;
 };
 
 MODULE_REGISTER_NS(Point,       TPoint<FIMPL>,        MSink);
@@ -77,6 +82,7 @@ MODULE_REGISTER_NS(ScalarPoint, TPoint<ScalarImplCR>, MSink);
 template <typename FImpl>
 TPoint<FImpl>::TPoint(const std::string name)
 : Module<PointPar>(name)
+, momphName_ (name + "_momph")
 {}
 
 // dependencies/products ///////////////////////////////////////////////////////
@@ -100,30 +106,37 @@ std::vector<std::string> TPoint<FImpl>::getOutput(void)
 template <typename FImpl>
 void TPoint<FImpl>::setup(void)
 {
-    unsigned int size;
+    envTmpLat(LatticeComplex, "coor");
-    
+    envCacheLat(LatticeComplex, momphName_);
-    size = env().template lattice4dSize<LatticeComplex>();
+    envCreate(SinkFn, getName(), 1, nullptr);
-    env().registerObject(getName(), size);
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 template <typename FImpl>
 void TPoint<FImpl>::execute(void)
-{
+{   
-    std::vector<Real> p = strToVec<Real>(par().mom);
-    LatticeComplex    ph(env().getGrid()), coor(env().getGrid());
-    Complex           i(0.0,1.0);
-    
     LOG(Message) << "Setting up point sink function for momentum ["
                  << par().mom << "]" << std::endl;
-    ph = zero;
+
-    for(unsigned int mu = 0; mu < env().getNd(); mu++)
+    auto &ph = envGet(LatticeComplex, momphName_);
+    
+    if (!hasPhase_)
     {
-        LatticeCoordinate(coor, mu);
+        Complex           i(0.0,1.0);
-        ph = ph + (p[mu]/env().getGrid()->_fdimensions[mu])*coor;
+        std::vector<Real> p;
+
+        envGetTmp(LatticeComplex, coor);
+        p  = strToVec<Real>(par().mom);
+        ph = zero;
+        for(unsigned int mu = 0; mu < env().getNd(); mu++)
+        {
+            LatticeCoordinate(coor, mu);
+            ph = ph + (p[mu]/env().getGrid()->_fdimensions[mu])*coor;
+        }
+        ph = exp((Real)(2*M_PI)*i*ph);
+        hasPhase_ = true;
     }
-    ph = exp((Real)(2*M_PI)*i*ph);
+    auto sink = [&ph](const PropagatorField &field)
-    auto sink = [ph](const PropagatorField &field)
     {
         SlicedPropagator res;
         PropagatorField  tmp = ph*field;
@@ -132,7 +145,7 @@ void TPoint<FImpl>::execute(void)
         
         return res;
     };
-    env().setObject(getName(), new SinkFn(sink));
+    envGet(SinkFn, getName()) = sink;
 }
 
 END_MODULE_NAMESPACE

extras/Hadrons/Modules/MSink/Smear.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MSink/Smear.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -62,6 +63,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -100,9 +102,7 @@ std::vector<std::string> TSmear<FImpl>::getOutput(void)
 template <typename FImpl>
 void TSmear<FImpl>::setup(void)
 {
-    unsigned int nt = env().getDim(Tp);
+    envCreate(SlicedPropagator, getName(), 1, env().getDim(Tp));
-    unsigned int size = nt * sizeof(SitePropagator);
-    env().registerObject(getName(), size);
 }
 
 // execution ///////////////////////////////////////////////////////////////////
@@ -113,11 +113,11 @@ void TSmear<FImpl>::execute(void)
                  << "' using sink function '" << par().sink << "'."
                  << std::endl;
 
-    SinkFn          &sink = *env().template getObject<SinkFn>(par().sink);
+    auto &sink = envGet(SinkFn, par().sink);
-    PropagatorField &q    = *env().template getObject<PropagatorField>(par().q);
+    auto &q    = envGet(PropagatorField, par().q);
-    SlicedPropagator *out = new SlicedPropagator(env().getDim(Tp));
+    auto &out  = envGet(SlicedPropagator, getName());
-    *out  = sink(q);
+    
-    env().setObject(getName(), out);
+    out = sink(q);
 }
 
 END_MODULE_NAMESPACE

extras/Hadrons/Modules/MSolver/RBPrecCG.hpp

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MSolver/RBPrecCG.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -61,7 +60,9 @@ public:
     virtual ~TRBPrecCG(void) = default;
     // dependencies/products
     virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getReference(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -83,11 +84,19 @@ TRBPrecCG<FImpl>::TRBPrecCG(const std::string name)
 template <typename FImpl>
 std::vector<std::string> TRBPrecCG<FImpl>::getInput(void)
 {
-    std::vector<std::string> in = {par().action};
+    std::vector<std::string> in = {};
     
     return in;
 }
 
+template <typename FImpl>
+std::vector<std::string> TRBPrecCG<FImpl>::getReference(void)
+{
+    std::vector<std::string> ref = {par().action};
+    
+    return ref;
+}
+
 template <typename FImpl>
 std::vector<std::string> TRBPrecCG<FImpl>::getOutput(void)
 {
@@ -100,17 +109,12 @@ std::vector<std::string> TRBPrecCG<FImpl>::getOutput(void)
 template <typename FImpl>
 void TRBPrecCG<FImpl>::setup(void)
 {
-    auto Ls = env().getObjectLs(par().action);
+    LOG(Message) << "setting up Schur red-black preconditioned CG for"
-    
+                 << " action '" << par().action << "' with residual "
-    env().registerObject(getName(), 0, Ls);
+                 << par().residual << std::endl;
-    env().addOwnership(getName(), par().action);
-}
 
-// execution ///////////////////////////////////////////////////////////////////
+    auto Ls     = env().getObjectLs(par().action);
-template <typename FImpl>
+    auto &mat   = envGet(FMat, par().action);
-void TRBPrecCG<FImpl>::execute(void)
-{
-    auto &mat   = *(env().template getObject<FMat>(par().action));
     auto solver = [&mat, this](FermionField &sol, const FermionField &source)
     {
         ConjugateGradient<FermionField>           cg(par().residual, 10000);
@@ -118,13 +122,14 @@ void TRBPrecCG<FImpl>::execute(void)
         
         schurSolver(mat, source, sol);
     };
-    
+    envCreate(SolverFn, getName(), Ls, solver);
-    LOG(Message) << "setting up Schur red-black preconditioned CG for"
-                 << " action '" << par().action << "' with residual "
-                 << par().residual << std::endl;
-    env().setObject(getName(), new SolverFn(solver));
 }
 
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TRBPrecCG<FImpl>::execute(void)
+{}
+
 END_MODULE_NAMESPACE
 
 END_HADRONS_NAMESPACE

extras/Hadrons/Modules/MSource/Point.hpp

@@ -4,10 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MSource/Point.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -72,6 +72,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -111,19 +112,20 @@ std::vector<std::string> TPoint<FImpl>::getOutput(void)
 template <typename FImpl>
 void TPoint<FImpl>::setup(void)
 {
-    env().template registerLattice<PropagatorField>(getName());
+    envCreateLat(PropagatorField, getName());
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 template <typename FImpl>
 void TPoint<FImpl>::execute(void)
 {
-    std::vector<int> position = strToVec<int>(par().position);
-    SitePropagator id;
-    
     LOG(Message) << "Creating point source at position [" << par().position
-                 << "]" << std::endl;
+                << "]" << std::endl;
-    PropagatorField &src = *env().template createLattice<PropagatorField>(getName());
+
+    std::vector<int> position = strToVec<int>(par().position);
+    auto             &src     = envGet(PropagatorField, getName());
+    SitePropagator   id;
+    
     id  = 1.;
     src = zero;
     pokeSite(id, src, position);

extras/Hadrons/Modules/MSource/SeqConserved.hpp

@@ -2,11 +2,12 @@
 
 Grid physics library, www.github.com/paboyle/Grid 
 
-Source file: extras/Hadrons/Modules/MContraction/SeqConserved.hpp
+Source file: extras/Hadrons/Modules/MSource/SeqConserved.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -83,6 +84,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -122,7 +124,7 @@ template <typename FImpl>
 void TSeqConserved<FImpl>::setup(void)
 {
     auto Ls_ = env().getObjectLs(par().action);
-    env().template registerLattice<PropagatorField>(getName(), Ls_);
+    envCreateLat(PropagatorField, getName(), Ls_);
 }
 
 // execution ///////////////////////////////////////////////////////////////////
@@ -142,9 +144,9 @@ void TSeqConserved<FImpl>::execute(void)
                      << par().mu << ") for " << par().tA << " <= t <= " 
                      << par().tB << std::endl;
     }
-    PropagatorField &src = *env().template createLattice<PropagatorField>(getName());
+    auto &src = envGet(PropagatorField, getName());
-    PropagatorField &q   = *env().template getObject<PropagatorField>(par().q);
+    auto &q   = envGet(PropagatorField, par().q);
-    FMat            &mat = *(env().template getObject<FMat>(par().action));
+    auto &mat = envGet(FMat, par().action);
 
     std::vector<Real> mom = strToVec<Real>(par().mom);
     mat.SeqConservedCurrent(q, src, par().curr_type, par().mu, 

extras/Hadrons/Modules/MSource/SeqGamma.hpp

@@ -4,11 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MSource/SeqGamma.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
-Copyright (C) 2017
 
 Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -81,10 +80,14 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
     virtual void execute(void);
+private:
+    bool        hasPhase_{false};
+    std::string momphName_, tName_;
 };
 
 MODULE_REGISTER_NS(SeqGamma, TSeqGamma<FIMPL>, MSource);
@@ -96,6 +99,8 @@ MODULE_REGISTER_NS(SeqGamma, TSeqGamma<FIMPL>, MSource);
 template <typename FImpl>
 TSeqGamma<FImpl>::TSeqGamma(const std::string name)
 : Module<SeqGammaPar>(name)
+, momphName_ (name + "_momph")
+, tName_ (name + "_t")
 {}
 
 // dependencies/products ///////////////////////////////////////////////////////
@@ -119,7 +124,10 @@ std::vector<std::string> TSeqGamma<FImpl>::getOutput(void)
 template <typename FImpl>
 void TSeqGamma<FImpl>::setup(void)
 {
-    env().template registerLattice<PropagatorField>(getName());
+    envCreateLat(PropagatorField, getName());
+    envCacheLat(Lattice<iScalar<vInteger>>, tName_);
+    envCacheLat(LatticeComplex, momphName_);
+    envTmpLat(LatticeComplex, "coor");
 }
 
 // execution ///////////////////////////////////////////////////////////////////
@@ -137,23 +145,29 @@ void TSeqGamma<FImpl>::execute(void)
                      << " sequential source for "
                      << par().tA << " <= t <= " << par().tB << std::endl;
     }
-    PropagatorField &src = *env().template createLattice<PropagatorField>(getName());
+    auto  &src = envGet(PropagatorField, getName());
-    PropagatorField &q   = *env().template getObject<PropagatorField>(par().q);
+    auto  &q   = envGet(PropagatorField, par().q);
-    Lattice<iScalar<vInteger>> t(env().getGrid());
+    auto  &ph  = envGet(LatticeComplex, momphName_);
-    LatticeComplex             ph(env().getGrid()), coor(env().getGrid());
+    auto  &t   = envGet(Lattice<iScalar<vInteger>>, tName_);
-    Gamma                      g(par().gamma);
+    Gamma g(par().gamma);
-    std::vector<Real>          p;
-    Complex                    i(0.0,1.0);
     
-    p  = strToVec<Real>(par().mom);
+    if (!hasPhase_)
-    ph = zero;
-    for(unsigned int mu = 0; mu < env().getNd(); mu++)
     {
-        LatticeCoordinate(coor, mu);
+        Complex           i(0.0,1.0);
-        ph = ph + p[mu]*coor*((1./(env().getGrid()->_fdimensions[mu])));
+        std::vector<Real> p;
+
+        envGetTmp(LatticeComplex, coor);
+        p  = strToVec<Real>(par().mom);
+        ph = zero;
+        for(unsigned int mu = 0; mu < env().getNd(); mu++)
+        {
+            LatticeCoordinate(coor, mu);
+            ph = ph + (p[mu]/env().getGrid()->_fdimensions[mu])*coor;
+        }
+        ph = exp((Real)(2*M_PI)*i*ph);
+        LatticeCoordinate(t, Tp);
+        hasPhase_ = true;
     }
-    ph = exp((Real)(2*M_PI)*i*ph);
-    LatticeCoordinate(t, Tp);
     src = where((t >= par().tA) and (t <= par().tB), ph*(g*q), 0.*q);
 }
 

extras/Hadrons/Modules/MSource/Wall.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MSource/Wall.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -73,10 +74,14 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
     virtual void execute(void);
+private:
+    bool        hasPhase_{false};
+    std::string momphName_, tName_;
 };
 
 MODULE_REGISTER_NS(Wall, TWall<FIMPL>, MSource);
@@ -88,13 +93,15 @@ MODULE_REGISTER_NS(Wall, TWall<FIMPL>, MSource);
 template <typename FImpl>
 TWall<FImpl>::TWall(const std::string name)
 : Module<WallPar>(name)
+, momphName_ (name + "_momph")
+, tName_ (name + "_t")
 {}
 
 // dependencies/products ///////////////////////////////////////////////////////
 template <typename FImpl>
 std::vector<std::string> TWall<FImpl>::getInput(void)
 {
-    std::vector<std::string> in;
+    std::vector<std::string> in = {};
     
     return in;
 }
@@ -111,7 +118,7 @@ std::vector<std::string> TWall<FImpl>::getOutput(void)
 template <typename FImpl>
 void TWall<FImpl>::setup(void)
 {
-    env().template registerLattice<PropagatorField>(getName());
+    envCreateLat(PropagatorField, getName());
 }
 
 // execution ///////////////////////////////////////////////////////////////////
@@ -121,21 +128,28 @@ void TWall<FImpl>::execute(void)
     LOG(Message) << "Generating wall source at t = " << par().tW 
                  << " with momentum " << par().mom << std::endl;
     
-    PropagatorField &src = *env().template createLattice<PropagatorField>(getName());
+    auto  &src = envGet(PropagatorField, getName());
-    Lattice<iScalar<vInteger>> t(env().getGrid());
+    auto  &ph  = envGet(LatticeComplex, momphName_);
-    LatticeComplex             ph(env().getGrid()), coor(env().getGrid());
+    auto  &t   = envGet(Lattice<iScalar<vInteger>>, tName_);
-    std::vector<Real>          p;
-    Complex                    i(0.0,1.0);
     
-    p  = strToVec<Real>(par().mom);
+    if (!hasPhase_)
-    ph = zero;
-    for(unsigned int mu = 0; mu < Nd; mu++)
     {
-        LatticeCoordinate(coor, mu);
+        Complex           i(0.0,1.0);
-        ph = ph + p[mu]*coor*((1./(env().getGrid()->_fdimensions[mu])));
+        std::vector<Real> p;
+
+        envGetTmp(LatticeComplex, coor);
+        p  = strToVec<Real>(par().mom);
+        ph = zero;
+        for(unsigned int mu = 0; mu < env().getNd(); mu++)
+        {
+            LatticeCoordinate(coor, mu);
+            ph = ph + (p[mu]/env().getGrid()->_fdimensions[mu])*coor;
+        }
+        ph = exp((Real)(2*M_PI)*i*ph);
+        LatticeCoordinate(t, Tp);
+        hasPhase_ = true;
     }
-    ph = exp((Real)(2*M_PI)*i*ph);
+
-    LatticeCoordinate(t, Tp);
     src = 1.;
     src = where((t == par().tW), src*ph, 0.*src);
 }

extras/Hadrons/Modules/MSource/Z2.hpp

@@ -4,8 +4,7 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MSource/Z2.hpp
 
-Copyright (C) 2015
+Copyright (C) 2015-2018
-Copyright (C) 2016
 
 Author: Antonin Portelli <antonin.portelli@me.com>
 
@@ -76,10 +75,14 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
     virtual void execute(void);
+private:
+    bool        hasT_{false};
+    std::string tName_;
 };
 
 MODULE_REGISTER_NS(Z2,       TZ2<FIMPL>,        MSource);
@@ -92,6 +95,7 @@ MODULE_REGISTER_NS(ScalarZ2, TZ2<ScalarImplCR>, MSource);
 template <typename FImpl>
 TZ2<FImpl>::TZ2(const std::string name)
 : Module<Z2Par>(name)
+, tName_ (name + "_t")
 {}
 
 // dependencies/products ///////////////////////////////////////////////////////
@@ -115,29 +119,36 @@ std::vector<std::string> TZ2<FImpl>::getOutput(void)
 template <typename FImpl>
 void TZ2<FImpl>::setup(void)
 {
-    env().template registerLattice<PropagatorField>(getName());
+    envCreateLat(PropagatorField, getName());
+    envCacheLat(Lattice<iScalar<vInteger>>, tName_);
+    envTmpLat(LatticeComplex, "eta");
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 template <typename FImpl>
 void TZ2<FImpl>::execute(void)
 {
-    Lattice<iScalar<vInteger>> t(env().getGrid());
-    LatticeComplex             eta(env().getGrid());
-    Complex                    shift(1., 1.);
-    
     if (par().tA == par().tB)
     {
         LOG(Message) << "Generating Z_2 wall source at t= " << par().tA
-        << std::endl;
+                     << std::endl;
     }
     else
     {
         LOG(Message) << "Generating Z_2 band for " << par().tA << " <= t <= "
-        << par().tB << std::endl;
+                     << par().tB << std::endl;
     }
-    PropagatorField &src = *env().template createLattice<PropagatorField>(getName());
+    
-    LatticeCoordinate(t, Tp);
+    auto    &src = envGet(PropagatorField, getName());
+    auto    &t   = envGet(Lattice<iScalar<vInteger>>, tName_);
+    Complex shift(1., 1.);
+
+    if (!hasT_)
+    {
+        LatticeCoordinate(t, Tp);
+        hasT_ = true;
+    }
+    envGetTmp(LatticeComplex, eta);
     bernoulli(*env().get4dRng(), eta);
     eta = (2.*eta - shift)*(1./::sqrt(2.));
     eta = where((t >= par().tA) and (t <= par().tB), eta, 0.*eta);

extras/Hadrons/Modules/MUtilities/TestSeqConserved.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MUtilities/TestSeqConserved.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -80,6 +81,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -121,38 +123,39 @@ void TTestSeqConserved<FImpl>::setup(void)
     auto Ls = env().getObjectLs(par().q);
     if (Ls != env().getObjectLs(par().action))
     {
-        HADRON_ERROR("Ls mismatch between quark action and propagator");
+        HADRON_ERROR(Size, "Ls mismatch between quark action and propagator");
     }
+    envTmpLat(PropagatorField, "tmp");
+    envTmpLat(LatticeComplex, "c");
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 template <typename FImpl>
 void TTestSeqConserved<FImpl>::execute(void)
 {
-    PropagatorField tmp(env().getGrid());
-    PropagatorField &q    = *env().template getObject<PropagatorField>(par().q);
-    PropagatorField &qSeq = *env().template getObject<PropagatorField>(par().qSeq);
-    FMat            &act  = *(env().template getObject<FMat>(par().action));
-    Gamma           g5(Gamma::Algebra::Gamma5);
-    Gamma::Algebra  gA = (par().curr == Current::Axial) ?
-                         Gamma::Algebra::Gamma5 :
-                         Gamma::Algebra::Identity;
-    Gamma           g(gA);
-    SitePropagator  qSite;
-    Complex         test_S, test_V, check_S, check_V;
-    std::vector<TComplex> check_buf;
-    LatticeComplex  c(env().getGrid());
-
     // Check sequential insertion of current gives same result as conserved 
     // current sink upon contraction. Assume q uses a point source.
-    std::vector<int> siteCoord;
+
+    auto                  &q    = envGet(PropagatorField, par().q);
+    auto                  &qSeq = envGet(PropagatorField, par().qSeq);
+    auto                  &act  = envGet(FMat, par().action);
+    Gamma                 g5(Gamma::Algebra::Gamma5);
+    Gamma::Algebra        gA = (par().curr == Current::Axial) ?
+                                  Gamma::Algebra::Gamma5 :
+                                  Gamma::Algebra::Identity;
+    Gamma                 g(gA);
+    SitePropagator        qSite;
+    Complex               test_S, test_V, check_S, check_V;
+    std::vector<TComplex> check_buf;
+    std::vector<int>      siteCoord;
+
+    envGetTmp(PropagatorField, tmp);
+    envGetTmp(LatticeComplex, c);
     siteCoord = strToVec<int>(par().origin);
     peekSite(qSite, qSeq, siteCoord);
     test_S = trace(qSite*g);
     test_V = trace(qSite*g*Gamma::gmu[par().mu]);
-
     act.ContractConservedCurrent(q, q, tmp, par().curr, par().mu);
-
     c = trace(tmp*g);
     sliceSum(c, check_buf, Tp);
     check_S = TensorRemove(check_buf[par().t_J]);

extras/Hadrons/Modules/MUtilities/TestSeqGamma.hpp

@@ -4,9 +4,10 @@ Grid physics library, www.github.com/paboyle/Grid
 
 Source file: extras/Hadrons/Modules/MUtilities/TestSeqGamma.hpp
 
-Copyright (C) 2017
+Copyright (C) 2015-2018
 
-Author: Andrew Lawson    <andrew.lawson1991@gmail.com>
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: Lanny91 <andrew.lawson@gmail.com>
 
 This program is free software; you can redistribute it and/or modify
 it under the terms of the GNU General Public License as published by
@@ -64,6 +65,7 @@ public:
     // dependency relation
     virtual std::vector<std::string> getInput(void);
     virtual std::vector<std::string> getOutput(void);
+protected:
     // setup
     virtual void setup(void);
     // execution
@@ -102,26 +104,27 @@ std::vector<std::string> TTestSeqGamma<FImpl>::getOutput(void)
 template <typename FImpl>
 void TTestSeqGamma<FImpl>::setup(void)
 {
-    
+    envTmpLat(LatticeComplex, "c");
 }
 
 // execution ///////////////////////////////////////////////////////////////////
 template <typename FImpl>
 void TTestSeqGamma<FImpl>::execute(void)
 {
-    PropagatorField &q    = *env().template getObject<PropagatorField>(par().q);
+    auto                  &q    = envGet(PropagatorField, par().q);
-    PropagatorField &qSeq = *env().template getObject<PropagatorField>(par().qSeq);
+    auto                  &qSeq = envGet(PropagatorField, par().qSeq);
-    LatticeComplex  c(env().getGrid());
+    Gamma                 g5(Gamma::Algebra::Gamma5);
-    Gamma           g5(Gamma::Algebra::Gamma5);
+    Gamma                 g(par().gamma);
-    Gamma           g(par().gamma);
+    SitePropagator        qSite;
-    SitePropagator  qSite;
+    Complex               test, check;
-    Complex         test, check;
     std::vector<TComplex> check_buf;
+    std::vector<int>      siteCoord;
 
     // Check sequential insertion of gamma matrix gives same result as 
     // insertion of gamma at sink upon contraction. Assume q uses a point 
     // source.
-    std::vector<int> siteCoord;
+    
+    envGetTmp(LatticeComplex, c);
     siteCoord = strToVec<int>(par().origin);
     peekSite(qSite, qSeq, siteCoord);
     test = trace(g*qSite);

extras/Hadrons/VirtualMachine.cc

@@ -0,0 +1,622 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/VirtualMachine.cc
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/Hadrons/VirtualMachine.hpp>
+#include <Grid/Hadrons/GeneticScheduler.hpp>
+#include <Grid/Hadrons/ModuleFactory.hpp>
+
+using namespace Grid;
+using namespace QCD;
+using namespace Hadrons;
+
+/******************************************************************************
+ *                      VirtualMachine implementation                         *
+ ******************************************************************************/
+// trajectory counter //////////////////////////////////////////////////////////
+void VirtualMachine::setTrajectory(const unsigned int traj)
+{
+    traj_ = traj;
+}
+
+unsigned int VirtualMachine::getTrajectory(void) const
+{
+    return traj_;
+}
+
+// module management ///////////////////////////////////////////////////////////
+void VirtualMachine::pushModule(VirtualMachine::ModPt &pt)
+{
+    std::string name = pt->getName();
+    
+    if (!hasModule(name))
+    {
+        std::vector<unsigned int> inputAddress;
+        unsigned int              address;
+        ModuleInfo                m;
+        
+        // module registration -------------------------------------------------
+        m.data = std::move(pt);
+        m.type = typeIdPt(*m.data.get());
+        m.name = name;
+        // input dependencies
+        for (auto &in: m.data->getInput())
+        {
+            if (!env().hasObject(in))
+            {
+                // if object does not exist, add it with no creator module
+                env().addObject(in , -1);
+            }
+            m.input.push_back(env().getObjectAddress(in));
+        }
+        // reference dependencies
+        for (auto &ref: m.data->getReference())
+        {
+            if (!env().hasObject(ref))
+            {
+                // if object does not exist, add it with no creator module
+                env().addObject(ref , -1);
+            }
+            m.input.push_back(env().getObjectAddress(ref));
+        }
+        auto inCopy = m.input;
+        // if module has inputs with references, they need to be added as
+        // an input
+        for (auto &in: inCopy)
+        {
+            int inm = env().getObjectModule(in);
+
+            if (inm > 0)
+            {
+                if (getModule(inm)->getReference().size() > 0)
+                {
+                    for (auto &rin: getModule(inm)->getReference())
+                    {
+                        m.input.push_back(env().getObjectAddress(rin));
+                    }
+                }
+            }
+        }
+        module_.push_back(std::move(m));
+        address              = static_cast<unsigned int>(module_.size() - 1);
+        moduleAddress_[name] = address;
+        // connecting outputs to potential inputs ------------------------------
+        for (auto &out: getModule(address)->getOutput())
+        {
+            if (!env().hasObject(out))
+            {
+                // output does not exists, add it
+                env().addObject(out, address);
+            }
+            else
+            {
+                if (env().getObjectModule(env().getObjectAddress(out)) < 0)
+                {
+                    // output exists but without creator, correct it
+                    env().setObjectModule(env().getObjectAddress(out), address);
+                }
+                else
+                {
+                    // output already fully registered, error
+                    HADRON_ERROR(Definition, "object '" + out
+                                 + "' is already produced by module '"
+                                 + module_[env().getObjectModule(out)].name
+                                 + "' (while pushing module '" + name + "')");
+                }
+                if (getModule(address)->getReference().size() > 0)
+                {
+                    // module has references, dependency should be propagated
+                    // to children modules; find module with `out` as an input
+                    // and add references to their input
+                    auto pred = [this, out](const ModuleInfo &n)
+                    {
+                        auto &in = n.input;
+                        auto it  = std::find(in.begin(), in.end(), 
+                                             env().getObjectAddress(out));
+                        
+                        return (it != in.end());
+                    };
+                    auto it = std::find_if(module_.begin(), module_.end(), pred);
+                    while (it != module_.end())
+                    {
+                        for (auto &ref: getModule(address)->getReference())
+                        {
+                            it->input.push_back(env().getObjectAddress(ref));
+                        }
+                        it = std::find_if(++it, module_.end(), pred);
+                    }   
+                }
+            }
+        }
+        graphOutdated_         = true;
+        memoryProfileOutdated_ = true;
+    }
+    else
+    {
+        HADRON_ERROR(Definition, "module '" + name + "' already exists");
+    }
+}
+
+unsigned int VirtualMachine::getNModule(void) const
+{
+    return module_.size();
+}
+
+void VirtualMachine::createModule(const std::string name, const std::string type,
+                                  XmlReader &reader)
+{
+    auto &factory = ModuleFactory::getInstance();
+    auto pt       = factory.create(type, name);
+    
+    pt->parseParameters(reader, "options");
+    pushModule(pt);
+}
+
+ModuleBase * VirtualMachine::getModule(const unsigned int address) const
+{
+    if (hasModule(address))
+    {
+        return module_[address].data.get();
+    }
+    else
+    {
+        HADRON_ERROR(Definition, "no module with address " + std::to_string(address));
+    }
+}
+
+ModuleBase * VirtualMachine::getModule(const std::string name) const
+{
+    return getModule(getModuleAddress(name));
+}
+
+unsigned int VirtualMachine::getModuleAddress(const std::string name) const
+{
+    if (hasModule(name))
+    {
+        return moduleAddress_.at(name);
+    }
+    else
+    {
+        HADRON_ERROR(Definition, "no module with name '" + name + "'");
+    }
+}
+
+std::string VirtualMachine::getModuleName(const unsigned int address) const
+{
+    if (hasModule(address))
+    {
+        return module_[address].name;
+    }
+    else
+    {
+        HADRON_ERROR(Definition, "no module with address " + std::to_string(address));
+    }
+}
+
+std::string VirtualMachine::getModuleType(const unsigned int address) const
+{
+    if (hasModule(address))
+    {
+        return typeName(module_[address].type);
+    }
+    else
+    {
+        HADRON_ERROR(Definition, "no module with address " + std::to_string(address));
+    }
+}
+
+std::string VirtualMachine::getModuleType(const std::string name) const
+{
+    return getModuleType(getModuleAddress(name));
+}
+
+std::string VirtualMachine::getModuleNamespace(const unsigned int address) const
+{
+    std::string type = getModuleType(address), ns;
+    
+    auto pos2 = type.rfind("::");
+    auto pos1 = type.rfind("::", pos2 - 2);
+    
+    return type.substr(pos1 + 2, pos2 - pos1 - 2);
+}
+
+std::string VirtualMachine::getModuleNamespace(const std::string name) const
+{
+    return getModuleNamespace(getModuleAddress(name));
+}
+
+bool VirtualMachine::hasModule(const unsigned int address) const
+{
+    return (address < module_.size());
+}
+
+bool VirtualMachine::hasModule(const std::string name) const
+{
+    return (moduleAddress_.find(name) != moduleAddress_.end());
+}
+
+// print VM content ////////////////////////////////////////////////////////////
+void VirtualMachine::printContent(void) const
+{
+    LOG(Debug) << "Modules: " << std::endl;
+    for (unsigned int i = 0; i < module_.size(); ++i)
+    {
+        LOG(Debug) << std::setw(4) << i << ": "
+                   << getModuleName(i) << std::endl;
+    }
+}
+
+// module graph ////////////////////////////////////////////////////////////////
+Graph<unsigned int> VirtualMachine::getModuleGraph(void)
+{
+    if (graphOutdated_)
+    {
+        makeModuleGraph();
+        graphOutdated_ = false;
+    }
+
+    return graph_;
+}
+
+void VirtualMachine::makeModuleGraph(void)
+{
+    Graph<unsigned int> graph;
+    
+    // create vertices
+    for (unsigned int m = 0; m < module_.size(); ++m)
+    {
+        graph.addVertex(m);
+    }
+    // create edges
+    for (unsigned int m = 0; m < module_.size(); ++m)
+    {
+        for (auto &in: module_[m].input)
+        {
+            graph.addEdge(env().getObjectModule(in), m);
+        }
+    }
+    graph_ = graph;
+}
+
+// memory profile //////////////////////////////////////////////////////////////
+const VirtualMachine::MemoryProfile & VirtualMachine::getMemoryProfile(void)
+{
+    if (memoryProfileOutdated_)
+    {
+        makeMemoryProfile();
+        memoryProfileOutdated_ = false;
+    }
+
+    return profile_;
+}
+
+void VirtualMachine::makeMemoryProfile(void)
+{
+    bool protect = env().objectsProtected();
+    bool hmsg    = HadronsLogMessage.isActive();
+    bool gmsg    = GridLogMessage.isActive();
+    bool err     = HadronsLogError.isActive();
+    auto program = getModuleGraph().topoSort();
+
+    resetProfile();
+    profile_.module.resize(getNModule());
+    env().protectObjects(false);
+    GridLogMessage.Active(false);
+    HadronsLogMessage.Active(false);
+    HadronsLogError.Active(false);
+    for (auto it = program.rbegin(); it != program.rend(); ++it) 
+    {
+        auto a = *it;
+
+        if (profile_.module[a].empty())
+        {
+            LOG(Debug) << "Profiling memory for module '" << module_[a].name
+                       << "' (" << a << ")..." << std::endl;
+            memoryProfile(a);
+            env().freeAll();
+        }
+    }
+    env().protectObjects(protect);
+    GridLogMessage.Active(gmsg);
+    HadronsLogMessage.Active(hmsg);
+    HadronsLogError.Active(err);
+    LOG(Debug) << "Memory profile:" << std::endl;
+    LOG(Debug) << "----------------" << std::endl;
+    for (unsigned int a = 0; a < profile_.module.size(); ++a)
+    {
+        LOG(Debug) << getModuleName(a) << " (" << a << ")" << std::endl;
+        for (auto &o: profile_.module[a])
+        {
+            LOG(Debug) << "|__ " << env().getObjectName(o.first) << " ("
+                       << sizeString(o.second) << ")" << std::endl;
+        }
+        LOG(Debug) << std::endl;
+    }
+    LOG(Debug) << "----------------" << std::endl;
+}
+
+void VirtualMachine::resetProfile(void)
+{
+    profile_.module.clear();
+    profile_.object.clear();
+}
+
+void VirtualMachine::resizeProfile(void)
+{
+    if (env().getMaxAddress() > profile_.object.size())
+    {
+        MemoryPrint empty;
+
+        empty.size   = 0;
+        empty.module = -1;
+        profile_.object.resize(env().getMaxAddress(), empty);
+    }
+}
+
+void VirtualMachine::updateProfile(const unsigned int address)
+{
+    resizeProfile();
+    for (unsigned int a = 0; a < env().getMaxAddress(); ++a)
+    {
+        if (env().hasCreatedObject(a) and (profile_.object[a].module == -1))
+        {
+            profile_.object[a].size     = env().getObjectSize(a);
+            profile_.object[a].storage  = env().getObjectStorage(a);
+            profile_.object[a].module   = address;
+            profile_.module[address][a] = profile_.object[a].size;
+            if (env().getObjectModule(a) < 0)
+            {
+                env().setObjectModule(a, address);
+            }
+        }
+    }
+}
+
+void VirtualMachine::cleanEnvironment(void)
+{
+    resizeProfile();
+    for (unsigned int a = 0; a < env().getMaxAddress(); ++a)
+    {
+        if (env().hasCreatedObject(a) and (profile_.object[a].module == -1))
+        {
+            env().freeObject(a);
+        }
+    }
+}
+
+void VirtualMachine::memoryProfile(const unsigned int address)
+{
+    auto m = getModule(address);
+
+    LOG(Debug) << "Setting up module '" << m->getName() 
+               << "' (" << address << ")..." << std::endl;
+    try
+    {
+        m->setup();
+        updateProfile(address);
+    }
+    catch (Exceptions::Definition &)
+    {
+        cleanEnvironment();
+        for (auto &in: m->getInput())
+        {
+            memoryProfile(env().getObjectModule(in));
+        }
+        for (auto &ref: m->getReference())
+        {
+            memoryProfile(env().getObjectModule(ref));
+        }
+        m->setup();
+        updateProfile(address);
+    }
+}
+
+void VirtualMachine::memoryProfile(const std::string name)
+{
+    memoryProfile(getModuleAddress(name));
+}
+
+// garbage collector ///////////////////////////////////////////////////////////
+VirtualMachine::GarbageSchedule 
+VirtualMachine::makeGarbageSchedule(const Program &p) const
+{
+    GarbageSchedule freeProg;
+    
+    freeProg.resize(p.size());
+    for (unsigned int a = 0; a < env().getMaxAddress(); ++a)
+    {
+        if (env().getObjectStorage(a) == Environment::Storage::temporary)
+        {
+            auto it = std::find(p.begin(), p.end(), env().getObjectModule(a));
+
+            if (it != p.end())
+            {
+                freeProg[std::distance(p.begin(), it)].insert(a);
+            }
+        }
+        else if (env().getObjectStorage(a) == Environment::Storage::object)
+        {
+            auto pred = [a, this](const unsigned int b)
+            {
+                auto &in = module_[b].input;
+                auto it  = std::find(in.begin(), in.end(), a);
+                
+                return (it != in.end()) or (b == env().getObjectModule(a));
+            };
+            auto it = std::find_if(p.rbegin(), p.rend(), pred);
+            if (it != p.rend())
+            {
+                freeProg[std::distance(it, p.rend()) - 1].insert(a);
+            }
+        }
+    }
+
+    return freeProg;
+}
+
+// high-water memory function //////////////////////////////////////////////////
+VirtualMachine::Size VirtualMachine::memoryNeeded(const Program &p)
+{
+    const MemoryProfile &profile = getMemoryProfile();
+    GarbageSchedule     freep    = makeGarbageSchedule(p);
+    Size                current = 0, max = 0;
+
+    for (unsigned int i = 0; i < p.size(); ++i)
+    {
+        for (auto &o: profile.module[p[i]])
+        {
+            current += o.second;
+        }
+        max = std::max(current, max);
+        for (auto &o: freep[i])
+        {
+            current -= profile.object[o].size;
+        }
+    }
+
+    return max;
+}
+
+// genetic scheduler ///////////////////////////////////////////////////////////
+VirtualMachine::Program VirtualMachine::schedule(const GeneticPar &par)
+{
+    typedef GeneticScheduler<Size, unsigned int> Scheduler;
+
+    auto graph = getModuleGraph();
+
+    //constrained topological sort using a genetic algorithm
+    LOG(Message) << "Scheduling computation..." << std::endl;
+    LOG(Message) << "               #module= " << graph.size() << std::endl;
+    LOG(Message) << "       population size= " << par.popSize << std::endl;
+    LOG(Message) << "       max. generation= " << par.maxGen << std::endl;
+    LOG(Message) << "  max. cst. generation= " << par.maxCstGen << std::endl;
+    LOG(Message) << "         mutation rate= " << par.mutationRate << std::endl;
+    
+    unsigned int          k = 0, gen, prevPeak, nCstPeak = 0;
+    std::random_device    rd;
+    Scheduler::Parameters gpar;
+    
+    gpar.popSize      = par.popSize;
+    gpar.mutationRate = par.mutationRate;
+    gpar.seed         = rd();
+    CartesianCommunicator::BroadcastWorld(0, &(gpar.seed), sizeof(gpar.seed));
+    Scheduler::ObjFunc memPeak = [this](const Program &p)->Size
+    {
+        return memoryNeeded(p);
+    };
+    Scheduler scheduler(graph, memPeak, gpar);
+    gen = 0;
+    do
+    {
+        LOG(Debug) << "Generation " << gen << ":" << std::endl;
+        scheduler.nextGeneration();
+        if (gen != 0)
+        {
+            if (prevPeak == scheduler.getMinValue())
+            {
+                nCstPeak++;
+            }
+            else
+            {
+                nCstPeak = 0;
+            }
+        }
+        
+        prevPeak = scheduler.getMinValue();
+        if (gen % 10 == 0)
+        {
+            LOG(Iterative) << "Generation " << gen << ": "
+                           << sizeString(scheduler.getMinValue()) << std::endl;
+        }
+        
+        gen++;
+    } while ((gen < par.maxGen) and (nCstPeak < par.maxCstGen));
+    
+    return scheduler.getMinSchedule();
+}
+
+// general execution ///////////////////////////////////////////////////////////
+#define BIG_SEP "==============="
+#define SEP     "---------------"
+#define MEM_MSG(size) sizeString(size)
+
+void VirtualMachine::executeProgram(const Program &p) const
+{
+    Size            memPeak = 0, sizeBefore, sizeAfter;
+    GarbageSchedule freeProg;
+    
+    // build garbage collection schedule
+    LOG(Debug) << "Building garbage collection schedule..." << std::endl;
+    freeProg = makeGarbageSchedule(p);
+
+    // program execution
+    LOG(Debug) << "Executing program..." << std::endl;
+    for (unsigned int i = 0; i < p.size(); ++i)
+    {
+        // execute module
+        LOG(Message) << SEP << " Measurement step " << i + 1 << "/"
+                     << p.size() << " (module '" << module_[p[i]].name
+                     << "') " << SEP << std::endl;
+        (*module_[p[i]].data)();
+        sizeBefore = env().getTotalSize();
+        // print used memory after execution
+        LOG(Message) << "Allocated objects: " << MEM_MSG(sizeBefore)
+                     << std::endl;
+        if (sizeBefore > memPeak)
+        {
+            memPeak = sizeBefore;
+        }
+        // garbage collection for step i
+        LOG(Message) << "Garbage collection..." << std::endl;
+        for (auto &j: freeProg[i])
+        {
+            env().freeObject(j);
+        }
+        // print used memory after garbage collection if necessary
+        sizeAfter = env().getTotalSize();
+        if (sizeBefore != sizeAfter)
+        {
+            LOG(Message) << "Allocated objects: " << MEM_MSG(sizeAfter)
+                            << std::endl;
+        }
+        else
+        {
+            LOG(Message) << "Nothing to free" << std::endl;
+        }
+    }
+}
+
+void VirtualMachine::executeProgram(const std::vector<std::string> &p) const
+{
+    Program pAddress;
+    
+    for (auto &n: p)
+    {
+        pAddress.push_back(getModuleAddress(n));
+    }
+    executeProgram(pAddress);
+}

extras/Hadrons/VirtualMachine.hpp

@@ -0,0 +1,207 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: extras/Hadrons/VirtualMachine.hpp
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+
+This program is free software; you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation; either version 2 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along
+with this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#ifndef Hadrons_VirtualMachine_hpp_
+#define Hadrons_VirtualMachine_hpp_
+
+#include <Grid/Hadrons/Global.hpp>
+#include <Grid/Hadrons/Graph.hpp>
+#include <Grid/Hadrons/Environment.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+#define DEFINE_VM_ALIAS \
+inline VirtualMachine & vm(void) const\
+{\
+    return VirtualMachine::getInstance();\
+}
+
+/******************************************************************************
+ *                   Virtual machine for module execution                     *
+ ******************************************************************************/
+// forward declaration of Module
+class ModuleBase;
+
+class VirtualMachine
+{
+    SINGLETON_DEFCTOR(VirtualMachine);
+public:
+    typedef SITE_SIZE_TYPE                      Size;
+    typedef std::unique_ptr<ModuleBase>         ModPt;
+    typedef std::vector<std::set<unsigned int>> GarbageSchedule;
+    typedef std::vector<unsigned int>           Program;
+    struct MemoryPrint
+    {
+        Size                 size;
+        Environment::Storage storage;
+        int                  module;
+    };
+    struct MemoryProfile
+    {
+        std::vector<std::map<unsigned int, Size>> module;
+        std::vector<MemoryPrint>                  object;
+    };
+    class GeneticPar: Serializable
+    {
+    public:
+        GeneticPar(void):
+            popSize{20}, maxGen{1000}, maxCstGen{100}, mutationRate{.1} {};
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(GeneticPar,
+                                        unsigned int, popSize,
+                                        unsigned int, maxGen,
+                                        unsigned int, maxCstGen,
+                                        double      , mutationRate);
+    };
+private:
+    struct ModuleInfo
+    {
+        const std::type_info      *type{nullptr};
+        std::string               name;
+        ModPt                     data{nullptr};
+        std::vector<unsigned int> input;
+        size_t                    maxAllocated;
+    };
+public:
+    // trajectory counter
+    void                setTrajectory(const unsigned int traj);
+    unsigned int        getTrajectory(void) const;
+    // module management
+    void                pushModule(ModPt &pt);
+    template <typename M>
+    void                createModule(const std::string name);
+    template <typename M>
+    void                createModule(const std::string name,
+                                         const typename M::Par &par);
+    void                createModule(const std::string name,
+                                         const std::string type,
+                                         XmlReader &reader);
+    unsigned int        getNModule(void) const;
+    ModuleBase *        getModule(const unsigned int address) const;
+    ModuleBase *        getModule(const std::string name) const;
+    template <typename M>
+    M *                 getModule(const unsigned int address) const;
+    template <typename M>
+    M *                 getModule(const std::string name) const;
+    unsigned int        getModuleAddress(const std::string name) const;
+    std::string         getModuleName(const unsigned int address) const;
+    std::string         getModuleType(const unsigned int address) const;
+    std::string         getModuleType(const std::string name) const;
+    std::string         getModuleNamespace(const unsigned int address) const;
+    std::string         getModuleNamespace(const std::string name) const;
+    bool                hasModule(const unsigned int address) const;
+    bool                hasModule(const std::string name) const;
+    // print VM content
+    void                printContent(void) const;
+    // module graph (could be a const reference if topoSort was const)
+    Graph<unsigned int> getModuleGraph(void);
+    // memory profile
+    const MemoryProfile &getMemoryProfile(void);
+    // garbage collector
+    GarbageSchedule     makeGarbageSchedule(const Program &p) const;
+    // high-water memory function
+    Size                memoryNeeded(const Program &p);
+    // genetic scheduler
+    Program             schedule(const GeneticPar &par);
+    // general execution
+    void                executeProgram(const Program &p) const;
+    void                executeProgram(const std::vector<std::string> &p) const;
+private:
+    // environment shortcut
+    DEFINE_ENV_ALIAS;
+    // module graph
+    void makeModuleGraph(void);
+    // memory profile
+    void makeMemoryProfile(void);
+    void resetProfile(void);
+    void resizeProfile(void);
+    void updateProfile(const unsigned int address);
+    void cleanEnvironment(void);
+    void memoryProfile(const std::string name);
+    void memoryProfile(const unsigned int address);
+private:
+    // general
+    unsigned int                        traj_;
+    // module and related maps
+    std::vector<ModuleInfo>             module_;
+    std::map<std::string, unsigned int> moduleAddress_;
+    std::string                         currentModule_{""};
+    // module graph
+    bool                                graphOutdated_{true};
+    Graph<unsigned int>                 graph_;
+    // memory profile
+    bool                                memoryProfileOutdated_{true};
+    MemoryProfile                       profile_;
+};
+
+/******************************************************************************
+ *                   VirtualMachine template implementation                   *
+ ******************************************************************************/
+// module management ///////////////////////////////////////////////////////////
+template <typename M>
+void VirtualMachine::createModule(const std::string name)
+{
+    ModPt pt(new M(name));
+    
+    pushModule(pt);
+}
+
+template <typename M>
+void VirtualMachine::createModule(const std::string name,
+                               const typename M::Par &par)
+{
+    ModPt pt(new M(name));
+    
+    static_cast<M *>(pt.get())->setPar(par);
+    pushModule(pt);
+}
+
+template <typename M>
+M * VirtualMachine::getModule(const unsigned int address) const
+{
+    if (auto *pt = dynamic_cast<M *>(getModule(address)))
+    {
+        return pt;
+    }
+    else
+    {
+        HADRON_ERROR(Definition, "module '" + module_[address].name
+                     + "' does not have type " + typeid(M).name()
+                     + "(has type: " + getModuleType(address) + ")");
+    }
+}
+
+template <typename M>
+M * VirtualMachine::getModule(const std::string name) const
+{
+    return getModule<M>(getModuleAddress(name));
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_VirtualMachine_hpp_

extras/Hadrons/modules.inc

@@ -2,43 +2,47 @@ modules_cc =\
   Modules/MContraction/WeakHamiltonianEye.cc \
   Modules/MContraction/WeakNeutral4ptDisc.cc \
   Modules/MContraction/WeakHamiltonianNonEye.cc \
+  Modules/MGauge/Unit.cc \
+  Modules/MGauge/StochEm.cc \
+  Modules/MGauge/Random.cc \
   Modules/MScalar/FreeProp.cc \
   Modules/MScalar/ChargedProp.cc \
-  Modules/MGauge/Random.cc \
+  Modules/MIO/LoadNersc.cc
-  Modules/MGauge/StochEm.cc \
-  Modules/MGauge/Unit.cc \
-  Modules/MGauge/Load.cc
 
 modules_hpp =\
-  Modules/MSolver/RBPrecCG.hpp \
-  Modules/MContraction/WardIdentity.hpp \
-  Modules/MContraction/Meson.hpp \
-  Modules/MContraction/Gamma3pt.hpp \
-  Modules/MContraction/DiscLoop.hpp \
-  Modules/MContraction/WeakHamiltonianEye.hpp \
   Modules/MContraction/Baryon.hpp \
+  Modules/MContraction/Meson.hpp \
   Modules/MContraction/WeakHamiltonian.hpp \
-  Modules/MContraction/WeakNeutral4ptDisc.hpp \
   Modules/MContraction/WeakHamiltonianNonEye.hpp \
+  Modules/MContraction/DiscLoop.hpp \
+  Modules/MContraction/WeakNeutral4ptDisc.hpp \
+  Modules/MContraction/Gamma3pt.hpp \
+  Modules/MContraction/WardIdentity.hpp \
+  Modules/MContraction/WeakHamiltonianEye.hpp \
+  Modules/MFermion/GaugeProp.hpp \
+  Modules/MSource/SeqConservedSummed.hpp \
+  Modules/MSource/SeqGamma.hpp \
+  Modules/MSource/Point.hpp \
+  Modules/MSource/Wall.hpp \
+  Modules/MSource/Z2.hpp \
+  Modules/MSource/SeqConserved.hpp \
+  Modules/MSink/Smear.hpp \
+  Modules/MSink/Point.hpp \
+  Modules/MSolver/RBPrecCG.hpp \
+  Modules/MGauge/Unit.hpp \
+  Modules/MGauge/Random.hpp \
+  Modules/MGauge/StochEm.hpp \
   Modules/MUtilities/TestSeqGamma.hpp \
   Modules/MUtilities/TestSeqConserved.hpp \
   Modules/MLoop/NoiseLoop.hpp \
   Modules/MScalar/FreeProp.hpp \
-  Modules/MScalar/ChargedProp.hpp \
   Modules/MScalar/Scalar.hpp \
-  Modules/MSink/Point.hpp \
+  Modules/MScalar/ChargedProp.hpp \
-  Modules/MSink/Smear.hpp \
-  Modules/MFermion/GaugeProp.hpp \
-  Modules/MSource/SeqConservedSummed.hpp \
-  Modules/MSource/Wall.hpp \
-  Modules/MSource/Point.hpp \
-  Modules/MSource/SeqConserved.hpp \
-  Modules/MSource/SeqGamma.hpp \
-  Modules/MSource/Z2.hpp \
-  Modules/MGauge/Load.hpp \
-  Modules/MGauge/StochEm.hpp \
-  Modules/MGauge/Random.hpp \
-  Modules/MGauge/Unit.hpp \
   Modules/MAction/DWF.hpp \
-  Modules/MAction/Wilson.hpp
+  Modules/MAction/Wilson.hpp \
+  Modules/MScalarSUN/TrMag.hpp \
+  Modules/MScalarSUN/TwoPoint.hpp \
+  Modules/MScalarSUN/TrPhi.hpp \
+  Modules/MIO/LoadNersc.hpp \
+Modules/MIO/LoadBinary.hpp
 

lib/Makefile.am

@@ -1,28 +1,18 @@
 extra_sources=
 extra_headers=
-if BUILD_COMMS_MPI
-  extra_sources+=communicator/Communicator_mpi.cc
-  extra_sources+=communicator/Communicator_base.cc
-endif
 
 if BUILD_COMMS_MPI3
   extra_sources+=communicator/Communicator_mpi3.cc
   extra_sources+=communicator/Communicator_base.cc
-endif
+  extra_sources+=communicator/SharedMemoryMPI.cc
-
+  extra_sources+=communicator/SharedMemory.cc
-if BUILD_COMMS_MPIT
-  extra_sources+=communicator/Communicator_mpit.cc
-  extra_sources+=communicator/Communicator_base.cc
-endif
-
-if BUILD_COMMS_SHMEM
-  extra_sources+=communicator/Communicator_shmem.cc
-  extra_sources+=communicator/Communicator_base.cc
 endif
 
 if BUILD_COMMS_NONE
   extra_sources+=communicator/Communicator_none.cc
   extra_sources+=communicator/Communicator_base.cc
+  extra_sources+=communicator/SharedMemoryNone.cc
+  extra_sources+=communicator/SharedMemory.cc
 endif
 
 if BUILD_HDF5

lib/algorithms/CoarsenedMatrix.h

@@ -103,29 +103,32 @@ namespace Grid {
     GridBase *CoarseGrid;
     GridBase *FineGrid;
     std::vector<Lattice<Fobj> > subspace;
+    int checkerboard;
 
-    Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid) : 
+  Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) : 
-      CoarseGrid(_CoarseGrid),
+    CoarseGrid(_CoarseGrid),
       FineGrid(_FineGrid),
-      subspace(nbasis,_FineGrid)
+      subspace(nbasis,_FineGrid),
+      checkerboard(_checkerboard)
 	{
 	};
   
     void Orthogonalise(void){
       CoarseScalar InnerProd(CoarseGrid); 
+      std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
       blockOrthogonalise(InnerProd,subspace);
+      std::cout << GridLogMessage <<" Gramm-Schmidt pass 2"<<std::endl;
+      blockOrthogonalise(InnerProd,subspace);
+      //      std::cout << GridLogMessage <<" Gramm-Schmidt checking orthogonality"<<std::endl;
+      //      CheckOrthogonal();
     } 
     void CheckOrthogonal(void){
       CoarseVector iProj(CoarseGrid); 
       CoarseVector eProj(CoarseGrid); 
-      Lattice<CComplex> pokey(CoarseGrid);
-
-      
       for(int i=0;i<nbasis;i++){
 	blockProject(iProj,subspace[i],subspace);
-
 	eProj=zero; 
-	for(int ss=0;ss<CoarseGrid->oSites();ss++){
+	parallel_for(int ss=0;ss<CoarseGrid->oSites();ss++){
 	  eProj._odata[ss](i)=CComplex(1.0);
 	}
 	eProj=eProj - iProj;
@@ -137,6 +140,7 @@ namespace Grid {
       blockProject(CoarseVec,FineVec,subspace);
     }
     void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
+      FineVec.checkerboard = subspace[0].checkerboard;
       blockPromote(CoarseVec,FineVec,subspace);
     }
     void CreateSubspaceRandom(GridParallelRNG &RNG){
@@ -147,6 +151,7 @@ namespace Grid {
       Orthogonalise();
     }
 
+    /*
     virtual void CreateSubspaceLanczos(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) 
     {
       // Run a Lanczos with sloppy convergence
@@ -195,7 +200,7 @@ namespace Grid {
 	  std::cout << GridLogMessage <<"subspace["<<b<<"] = "<<norm2(subspace[b])<<std::endl;
 	}
     }
-
+    */
     virtual void CreateSubspace(GridParallelRNG  &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) {
 
       RealD scale;

lib/algorithms/LinearOperator.h

@@ -162,15 +162,10 @@ namespace Grid {
 	_Mat.M(in,out);
       }
       void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
-	ComplexD dot;
-
 	_Mat.M(in,out);
 	
-	dot= innerProduct(in,out);
+	ComplexD dot= innerProduct(in,out); n1=real(dot);
-	n1=real(dot);
+	n2=norm2(out);
-
-	dot = innerProduct(out,out);
-	n2=real(dot);
       }
       void HermOp(const Field &in, Field &out){
 	_Mat.M(in,out);
@@ -192,10 +187,10 @@ namespace Grid {
 	ni=Mpc(in,tmp);
 	no=MpcDag(tmp,out);
       }
-      void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+      virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
 	MpcDagMpc(in,out,n1,n2);
       }
-      void HermOp(const Field &in, Field &out){
+      virtual void HermOp(const Field &in, Field &out){
 	RealD n1,n2;
 	HermOpAndNorm(in,out,n1,n2);
       }
@@ -212,7 +207,6 @@ namespace Grid {
       void OpDir  (const Field &in, Field &out,int dir,int disp) {
 	assert(0);
       }
-
     };
     template<class Matrix,class Field>
       class SchurDiagMooeeOperator :  public SchurOperatorBase<Field> {
@@ -270,7 +264,6 @@ namespace Grid {
 	return axpy_norm(out,-1.0,tmp,in);
       }
     };
-
     template<class Matrix,class Field>
       class SchurDiagTwoOperator :  public SchurOperatorBase<Field> {
     protected:
@@ -299,6 +292,59 @@ namespace Grid {
 	return axpy_norm(out,-1.0,tmp,in);
       }
     };
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    // Left  handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) psi = eta  -->  ( 1 - Moo^-1 Moe Mee^-1 Meo ) psi = Moo^-1 eta
+    // Right handed Moo^-1 ; (Moo - Moe Mee^-1 Meo) Moo^-1 Moo psi = eta  -->  ( 1 - Moe Mee^-1 Meo ) Moo^-1 phi=eta ; psi = Moo^-1 phi
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    template<class Matrix,class Field> using SchurDiagOneRH = SchurDiagTwoOperator<Matrix,Field> ;
+    template<class Matrix,class Field> using SchurDiagOneLH = SchurDiagOneOperator<Matrix,Field> ;
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    //  Staggered use
+    ///////////////////////////////////////////////////////////////////////////////////////////////////
+    template<class Matrix,class Field>
+      class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
+    protected:
+      Matrix &_Mat;
+    public:
+      SchurStaggeredOperator (Matrix &Mat): _Mat(Mat){};
+      virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
+	GridLogIterative.TimingMode(1);
+	std::cout << GridLogIterative << " HermOpAndNorm "<<std::endl;
+	n2 = Mpc(in,out);
+	std::cout << GridLogIterative << " HermOpAndNorm.Mpc "<<std::endl;
+	ComplexD dot= innerProduct(in,out);
+	std::cout << GridLogIterative << " HermOpAndNorm.innerProduct "<<std::endl;
+	n1 = real(dot);
+      }
+      virtual void HermOp(const Field &in, Field &out){
+	std::cout << GridLogIterative << " HermOp "<<std::endl;
+	Mpc(in,out);
+      }
+      virtual  RealD Mpc      (const Field &in, Field &out) {
+	Field tmp(in._grid);
+	Field tmp2(in._grid);
+
+	std::cout << GridLogIterative << " HermOp.Mpc "<<std::endl;
+	_Mat.Mooee(in,out);
+	_Mat.Mooee(out,tmp);
+	std::cout << GridLogIterative << " HermOp.MooeeMooee "<<std::endl;
+
+	_Mat.Meooe(in,out);
+	_Mat.Meooe(out,tmp2);
+	std::cout << GridLogIterative << " HermOp.MeooeMeooe "<<std::endl;
+
+	RealD nn=axpy_norm(out,-1.0,tmp2,tmp);
+	std::cout << GridLogIterative << " HermOp.axpy_norm "<<std::endl;
+	return nn;
+      }
+      virtual  RealD MpcDag   (const Field &in, Field &out){
+	return Mpc(in,out);
+      }
+      virtual void MpcDagMpc(const Field &in, Field &out,RealD &ni,RealD &no) {
+	assert(0);// Never need with staggered
+      }
+    };
+    template<class Matrix,class Field> using SchurStagOperator = SchurStaggeredOperator<Matrix,Field>;
 
 
     /////////////////////////////////////////////////////////////
@@ -314,6 +360,14 @@ namespace Grid {
       virtual void operator() (const Field &in, Field &out) = 0;
     };
 
+    template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {
+    public:
+      void operator() (const Field &in, Field &out){
+	out = in;
+      };
+    };
+
+
     /////////////////////////////////////////////////////////////
     // Base classes for Multishift solvers for operators
     /////////////////////////////////////////////////////////////
@@ -336,6 +390,64 @@ namespace Grid {
      };
     */
 
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // Hermitian operator Linear function and operator function
+  ////////////////////////////////////////////////////////////////////////////////////////////
+    template<class Field>
+      class HermOpOperatorFunction : public OperatorFunction<Field> {
+      void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
+	Linop.HermOp(in,out);
+      };
+    };
+
+    template<typename Field>
+      class PlainHermOp : public LinearFunction<Field> {
+    public:
+      LinearOperatorBase<Field> &_Linop;
+      
+      PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop) 
+      {}
+      
+      void operator()(const Field& in, Field& out) {
+	_Linop.HermOp(in,out);
+      }
+    };
+
+    template<typename Field>
+    class FunctionHermOp : public LinearFunction<Field> {
+    public:
+      OperatorFunction<Field>   & _poly;
+      LinearOperatorBase<Field> &_Linop;
+      
+      FunctionHermOp(OperatorFunction<Field> & poly,LinearOperatorBase<Field>& linop) 
+	: _poly(poly), _Linop(linop) {};
+      
+      void operator()(const Field& in, Field& out) {
+	_poly(_Linop,in,out);
+      }
+    };
+
+  template<class Field>
+  class Polynomial : public OperatorFunction<Field> {
+  private:
+    std::vector<RealD> Coeffs;
+  public:
+    Polynomial(std::vector<RealD> &_Coeffs) : Coeffs(_Coeffs) { };
+
+    // Implement the required interface
+    void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
+
+      Field AtoN(in._grid);
+      Field Mtmp(in._grid);
+      AtoN = in;
+      out = AtoN*Coeffs[0];
+      for(int n=1;n<Coeffs.size();n++){
+	Mtmp = AtoN;
+	Linop.HermOp(Mtmp,AtoN);
+	out=out+AtoN*Coeffs[n];
+      }
+    };
+  };
 
 }
 

lib/algorithms/approx/Chebyshev.h

@@ -8,6 +8,7 @@
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 Author: paboyle <paboyle@ph.ed.ac.uk>
+Author: Christoph Lehner <clehner@bnl.gov>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -33,41 +34,12 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 
 namespace Grid {
 
-  ////////////////////////////////////////////////////////////////////////////////////////////
+struct ChebyParams : Serializable {
-  // Simple general polynomial with user supplied coefficients
+  GRID_SERIALIZABLE_CLASS_MEMBERS(ChebyParams,
-  ////////////////////////////////////////////////////////////////////////////////////////////
+				  RealD, alpha,  
-  template<class Field>
+				  RealD, beta,   
-  class HermOpOperatorFunction : public OperatorFunction<Field> {
+				  int, Npoly);
-    void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
+};
-      Linop.HermOp(in,out);
-    };
-  };
-
-  template<class Field>
-  class Polynomial : public OperatorFunction<Field> {
-  private:
-    std::vector<RealD> Coeffs;
-  public:
-    Polynomial(std::vector<RealD> &_Coeffs) : Coeffs(_Coeffs) { };
-
-    // Implement the required interface
-    void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
-
-      Field AtoN(in._grid);
-      Field Mtmp(in._grid);
-      AtoN = in;
-      out = AtoN*Coeffs[0];
-//            std::cout <<"Poly in " <<norm2(in)<<" size "<< Coeffs.size()<<std::endl;
-//            std::cout <<"Coeffs[0]= "<<Coeffs[0]<< " 0 " <<norm2(out)<<std::endl;
-      for(int n=1;n<Coeffs.size();n++){
-	Mtmp = AtoN;
-	Linop.HermOp(Mtmp,AtoN);
-	out=out+AtoN*Coeffs[n];
-//            std::cout <<"Coeffs "<<n<<"= "<< Coeffs[n]<< " 0 " <<std::endl;
-//		std::cout << n<<" " <<norm2(out)<<std::endl;
-      }
-    };
-  };
 
   ////////////////////////////////////////////////////////////////////////////////////////////
   // Generic Chebyshev approximations
@@ -82,8 +54,10 @@ namespace Grid {
 
   public:
     void csv(std::ostream &out){
-	RealD diff = hi-lo;
+      RealD diff = hi-lo;
-      for (RealD x=lo-0.2*diff; x<hi+0.2*diff; x+=(hi-lo)/1000) {
+      RealD delta = (hi-lo)*1.0e-9;
+      for (RealD x=lo; x<hi; x+=delta) {
+	delta*=1.1;
 	RealD f = approx(x);
 	out<< x<<" "<<f<<std::endl;
       }
@@ -99,6 +73,7 @@ namespace Grid {
     };
 
     Chebyshev(){};
+    Chebyshev(ChebyParams p){ Init(p.alpha,p.beta,p.Npoly);};
     Chebyshev(RealD _lo,RealD _hi,int _order, RealD (* func)(RealD) ) {Init(_lo,_hi,_order,func);};
     Chebyshev(RealD _lo,RealD _hi,int _order) {Init(_lo,_hi,_order);};
 
@@ -193,6 +168,47 @@ namespace Grid {
       return sum;
     };
 
+    RealD approxD(RealD x)
+    {
+      RealD Un;
+      RealD Unm;
+      RealD Unp;
+      
+      RealD y=( x-0.5*(hi+lo))/(0.5*(hi-lo));
+      
+      RealD U0=1;
+      RealD U1=2*y;
+      
+      RealD sum;
+      sum = Coeffs[1]*U0;
+      sum+= Coeffs[2]*U1*2.0;
+      
+      Un =U1;
+      Unm=U0;
+      for(int i=2;i<order-1;i++){
+	Unp=2*y*Un-Unm;
+	Unm=Un;
+	Un =Unp;
+	sum+= Un*Coeffs[i+1]*(i+1.0);
+      }
+      return sum/(0.5*(hi-lo));
+    };
+    
+    RealD approxInv(RealD z, RealD x0, int maxiter, RealD resid) {
+      RealD x = x0;
+      RealD eps;
+      
+      int i;
+      for (i=0;i<maxiter;i++) {
+	eps = approx(x) - z;
+	if (fabs(eps / z) < resid)
+	  return x;
+	x = x - eps / approxD(x);
+      }
+      
+      return std::numeric_limits<double>::quiet_NaN();
+    }
+    
     // Implement the required interface
     void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
 

lib/algorithms/iterative/ConjugateGradient.h

@@ -78,12 +78,12 @@ class ConjugateGradient : public OperatorFunction<Field> {
     cp = a;
     ssq = norm2(src);
 
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: guess " << guess << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: guess " << guess << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:   src " << ssq << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:   src " << ssq << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:    mp " << d << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:    mp " << d << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:   mmp " << b << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:   mmp " << b << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:  cp,r " << cp << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:  cp,r " << cp << std::endl;
-    std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient:     p " << a << std::endl;
+    std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient:     p " << a << std::endl;
 
     RealD rsq = Tolerance * Tolerance * ssq;
 
@@ -92,7 +92,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
       return;
     }
 
-    std::cout << GridLogIterative << std::setprecision(4)
+    std::cout << GridLogIterative << std::setprecision(8)
               << "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
 
     GridStopWatch LinalgTimer;

lib/algorithms/iterative/ImplicitlyRestartedLanczos.h

@@ -7,8 +7,9 @@
     Copyright (C) 2015
 
 Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: Chulwoo Jung
+Author: paboyle <paboyle@ph.ed.ac.uk>
-Author: Guido Cossu
+Author: Chulwoo Jung <chulwoo@bnl.gov>
+Author: Christoph Lehner <clehner@bnl.gov>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
@@ -27,125 +28,288 @@ Author: Guido Cossu
     See the full license in the file "LICENSE" in the top level distribution directory
     *************************************************************************************/
     /*  END LEGAL */
-#ifndef GRID_IRL_H
+#ifndef GRID_BIRL_H
-#define GRID_IRL_H
+#define GRID_BIRL_H
 
 #include <string.h> //memset
+//#include <zlib.h>
+#include <sys/stat.h>
 
-namespace Grid {
+namespace Grid { 
 
-  enum IRLdiagonalisation { 
+  ////////////////////////////////////////////////////////
-    IRLdiagonaliseWithDSTEGR,
+  // Move following 100 LOC to lattice/Lattice_basis.h
-    IRLdiagonaliseWithQR,
+  ////////////////////////////////////////////////////////
-    IRLdiagonaliseWithEigen
-  };
-
-////////////////////////////////////////////////////////////////////////////////
-// Helper class for sorting the evalues AND evectors by Field
-// Use pointer swizzle on vectors
-////////////////////////////////////////////////////////////////////////////////
 template<class Field>
-class SortEigen {
+void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k) 
- private:
+{
-  static bool less_lmd(RealD left,RealD right){
+  for(int j=0; j<k; ++j){
-    return left > right;
+    auto ip = innerProduct(basis[j],w);
-  }  
+    w = w - ip*basis[j];
-  static bool less_pair(std::pair<RealD,Field const*>& left,
+  }
-                        std::pair<RealD,Field const*>& right){
+}
-    return left.first > (right.first);
-  }  
-  
- public:
-  void push(std::vector<RealD>& lmd,std::vector<Field>& evec,int N) {
-    
-    ////////////////////////////////////////////////////////////////////////
-    // PAB: FIXME: VERY VERY VERY wasteful: takes a copy of the entire vector set.
-    //    : The vector reorder should be done by pointer swizzle somehow
-    ////////////////////////////////////////////////////////////////////////
-    std::vector<Field> cpy(lmd.size(),evec[0]._grid);
-    for(int i=0;i<lmd.size();i++) cpy[i] = evec[i];
-    
-    std::vector<std::pair<RealD, Field const*> > emod(lmd.size());    
 
-    for(int i=0;i<lmd.size();++i)  emod[i] = std::pair<RealD,Field const*>(lmd[i],&cpy[i]);
+template<class Field>
-
+void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, int k0,int k1,int Nm) 
-    partial_sort(emod.begin(),emod.begin()+N,emod.end(),less_pair);
+{
-
+  typedef typename Field::vector_object vobj;
-    typename std::vector<std::pair<RealD, Field const*> >::iterator it = emod.begin();
+  GridBase* grid = basis[0]._grid;
-    for(int i=0;i<N;++i){
+      
-      lmd[i]=it->first;
+  parallel_region
-      evec[i]=*(it->second);
+  {
-      ++it;
+    std::vector < vobj > B(Nm); // Thread private
+        
+    parallel_for_internal(int ss=0;ss < grid->oSites();ss++){
+      for(int j=j0; j<j1; ++j) B[j]=0.;
+      
+      for(int j=j0; j<j1; ++j){
+	for(int k=k0; k<k1; ++k){
+	  B[j] +=Qt(j,k) * basis[k]._odata[ss];
+	}
+      }
+      for(int j=j0; j<j1; ++j){
+	  basis[j]._odata[ss] = B[j];
+      }
     }
   }
-  void push(std::vector<RealD>& lmd,int N) {
+}
-    std::partial_sort(lmd.begin(),lmd.begin()+N,lmd.end(),less_lmd);
+
+// Extract a single rotated vector
+template<class Field>
+void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm) 
+{
+  typedef typename Field::vector_object vobj;
+  GridBase* grid = basis[0]._grid;
+
+  result.checkerboard = basis[0].checkerboard;
+  parallel_for(int ss=0;ss < grid->oSites();ss++){
+    vobj B = zero;
+    for(int k=k0; k<k1; ++k){
+      B +=Qt(j,k) * basis[k]._odata[ss];
+    }
+    result._odata[ss] = B;
   }
-  bool saturated(RealD lmd, RealD thrs) {
+}
-    return fabs(lmd) > fabs(thrs);
+
+template<class Field>
+void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, std::vector<int>& idx) 
+{
+  int vlen = idx.size();
+
+  assert(vlen>=1);
+  assert(vlen<=sort_vals.size());
+  assert(vlen<=_v.size());
+
+  for (size_t i=0;i<vlen;i++) {
+
+    if (idx[i] != i) {
+
+      //////////////////////////////////////
+      // idx[i] is a table of desired sources giving a permutation.
+      // Swap v[i] with v[idx[i]].
+      // Find  j>i for which _vnew[j] = _vold[i],
+      // track the move idx[j] => idx[i]
+      // track the move idx[i] => i
+      //////////////////////////////////////
+      size_t j;
+      for (j=i;j<idx.size();j++)
+	if (idx[j]==i)
+	  break;
+
+      assert(idx[i] > i);     assert(j!=idx.size());      assert(idx[j]==i);
+
+      std::swap(_v[i]._odata,_v[idx[i]]._odata); // should use vector move constructor, no data copy
+      std::swap(sort_vals[i],sort_vals[idx[i]]);
+
+      idx[j] = idx[i];
+      idx[i] = i;
+    }
   }
-};
+}
+
+inline std::vector<int> basisSortGetIndex(std::vector<RealD>& sort_vals) 
+{
+  std::vector<int> idx(sort_vals.size());
+  std::iota(idx.begin(), idx.end(), 0);
+
+  // sort indexes based on comparing values in v
+  std::sort(idx.begin(), idx.end(), [&sort_vals](int i1, int i2) {
+    return ::fabs(sort_vals[i1]) < ::fabs(sort_vals[i2]);
+  });
+  return idx;
+}
+
+template<class Field>
+void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, bool reverse) 
+{
+  std::vector<int> idx = basisSortGetIndex(sort_vals);
+  if (reverse)
+    std::reverse(idx.begin(), idx.end());
+  
+  basisReorderInPlace(_v,sort_vals,idx);
+}
+
+// PAB: faster to compute the inner products first then fuse loops.
+// If performance critical can improve.
+template<class Field>
+void basisDeflate(const std::vector<Field> &_v,const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
+  result = zero;
+  assert(_v.size()==eval.size());
+  int N = (int)_v.size();
+  for (int i=0;i<N;i++) {
+    Field& tmp = _v[i];
+    axpy(result,TensorRemove(innerProduct(tmp,src_orig)) / eval[i],tmp,result);
+  }
+}
 
 /////////////////////////////////////////////////////////////
 // Implicitly restarted lanczos
 /////////////////////////////////////////////////////////////
+template<class Field> class ImplicitlyRestartedLanczosTester 
+{
+ public:
+  virtual int TestConvergence(int j,RealD resid,Field &evec, RealD &eval,RealD evalMaxApprox)=0;
+  virtual int ReconstructEval(int j,RealD resid,Field &evec, RealD &eval,RealD evalMaxApprox)=0;
+};
+
+enum IRLdiagonalisation { 
+  IRLdiagonaliseWithDSTEGR,
+  IRLdiagonaliseWithQR,
+  IRLdiagonaliseWithEigen
+};
+
+template<class Field> class ImplicitlyRestartedLanczosHermOpTester  : public ImplicitlyRestartedLanczosTester<Field>
+{
+ public:
+  LinearFunction<Field>       &_HermOp;
+  ImplicitlyRestartedLanczosHermOpTester(LinearFunction<Field> &HermOp) : _HermOp(HermOp)  {  };
+  int ReconstructEval(int j,RealD resid,Field &B, RealD &eval,RealD evalMaxApprox)
+  {
+    return TestConvergence(j,resid,B,eval,evalMaxApprox);
+  }
+  int TestConvergence(int j,RealD eresid,Field &B, RealD &eval,RealD evalMaxApprox)
+  {
+    Field v(B);
+    RealD eval_poly = eval;
+    // Apply operator
+    _HermOp(B,v);
+
+    RealD vnum = real(innerProduct(B,v)); // HermOp.
+    RealD vden = norm2(B);
+    RealD vv0  = norm2(v);
+    eval   = vnum/vden;
+    v -= eval*B;
+
+    RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
+
+    std::cout.precision(13);
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<std::endl;
+
+    int conv=0;
+    if( (vv<eresid*eresid) ) conv = 1;
+
+    return conv;
+  }
+};
+
 template<class Field> 
 class ImplicitlyRestartedLanczos {
-
+ private:
-private:       
+  const RealD small = 1.0e-8;
-
+  int MaxIter;
-  int MaxIter;   // Max iterations
+  int MinRestart; // Minimum number of restarts; only check for convergence after
-  int Nstop;     // Number of evecs checked for convergence
+  int Nstop;   // Number of evecs checked for convergence
-  int Nk;        // Number of converged sought
+  int Nk;      // Number of converged sought
-  int Nm;        // Nm -- total number of vectors
+  //  int Np;      // Np -- Number of spare vecs in krylov space //  == Nm - Nk
-  RealD eresid;
+  int Nm;      // Nm -- total number of vectors
   IRLdiagonalisation diagonalisation;
-  ////////////////////////////////////
+  int orth_period;
+    
+  RealD OrthoTime;
+  RealD eresid, betastp;
+  ////////////////////////////////
   // Embedded objects
-  ////////////////////////////////////
+  ////////////////////////////////
-           SortEigen<Field> _sort;
+  LinearFunction<Field>       &_PolyOp;
-  LinearOperatorBase<Field> &_Linop;
+  LinearFunction<Field>       &_HermOp;
-    OperatorFunction<Field> &_poly;
+  ImplicitlyRestartedLanczosTester<Field> &_Tester;
-
+  // Default tester provided (we need a ref to something in default case)
+  ImplicitlyRestartedLanczosHermOpTester<Field> SimpleTester;
   /////////////////////////
   // Constructor
   /////////////////////////
+  
 public:       
- ImplicitlyRestartedLanczos(LinearOperatorBase<Field> &Linop, // op
+  //////////////////////////////////////////////////////////////////
-			    OperatorFunction<Field> & poly,   // polynomial
+  // PAB:
-			    int _Nstop, // really sought vecs
+  //////////////////////////////////////////////////////////////////
-			    int _Nk,    // sought vecs
+  // Too many options  & knobs. 
-			    int _Nm,    // total vecs
+  // Eliminate:
-			    RealD _eresid, // resid in lmd deficit 
+  //   orth_period
-			    int _MaxIter,  // Max iterations
+  //   betastp
-			    IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen ) :
+  //   MinRestart
-    _Linop(Linop),    _poly(poly),
+  //
-      Nstop(_Nstop), Nk(_Nk), Nm(_Nm),
+  // Do we really need orth_period
-      eresid(_eresid),  MaxIter(_MaxIter),
+  // What is the theoretical basis & guarantees of betastp ?
-      diagonalisation(_diagonalisation)
+  // Nstop=Nk viable?
-      { };
+  // MinRestart avoidable with new convergence test?
+  // Could cut to PolyOp, HermOp, Tester, Nk, Nm, resid, maxiter (+diagonalisation)
+  // HermOp could be eliminated if we dropped the Power method for max eval.
+  // -- also: The eval, eval2, eval2_copy stuff is still unnecessarily unclear
+  //////////////////////////////////////////////////////////////////
+ ImplicitlyRestartedLanczos(LinearFunction<Field> & PolyOp,
+			    LinearFunction<Field> & HermOp,
+			    ImplicitlyRestartedLanczosTester<Field> & Tester,
+			    int _Nstop, // sought vecs
+			    int _Nk, // sought vecs
+			    int _Nm, // spare vecs
+			    RealD _eresid, // resid in lmdue deficit 
+			    int _MaxIter, // Max iterations
+			    RealD _betastp=0.0, // if beta(k) < betastp: converged
+			    int _MinRestart=1, int _orth_period = 1,
+			    IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
+    SimpleTester(HermOp), _PolyOp(PolyOp),      _HermOp(HermOp), _Tester(Tester),
+    Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
+    eresid(_eresid),      betastp(_betastp),
+    MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
+    orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
+
+    ImplicitlyRestartedLanczos(LinearFunction<Field> & PolyOp,
+			       LinearFunction<Field> & HermOp,
+			       int _Nstop, // sought vecs
+			       int _Nk, // sought vecs
+			       int _Nm, // spare vecs
+			       RealD _eresid, // resid in lmdue deficit 
+			       int _MaxIter, // Max iterations
+			       RealD _betastp=0.0, // if beta(k) < betastp: converged
+			       int _MinRestart=1, int _orth_period = 1,
+			       IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
+    SimpleTester(HermOp),  _PolyOp(PolyOp),      _HermOp(HermOp), _Tester(SimpleTester),
+    Nstop(_Nstop)  ,      Nk(_Nk),      Nm(_Nm),
+    eresid(_eresid),      betastp(_betastp),
+    MaxIter(_MaxIter)  ,      MinRestart(_MinRestart),
+    orth_period(_orth_period), diagonalisation(_diagonalisation)  { };
 
   ////////////////////////////////
   // Helpers
   ////////////////////////////////
-  static RealD normalise(Field& v) 
+  template<typename T>  static RealD normalise(T& v) 
   {
     RealD nn = norm2(v);
     nn = sqrt(nn);
     v = v * (1.0/nn);
     return nn;
   }
-  
+
-  void orthogonalize(Field& w, std::vector<Field>& evec, int k)
+  void orthogonalize(Field& w, std::vector<Field>& evec,int k)
   {
-    typedef typename Field::scalar_type MyComplex;
+    OrthoTime-=usecond()/1e6;
-    MyComplex ip;
+    basisOrthogonalize(evec,w,k);
-    
-    for(int j=0; j<k; ++j){
-      ip = innerProduct(evec[j],w); 
-      w = w - ip * evec[j];
-    }
     normalise(w);
+    OrthoTime+=usecond()/1e6;
   }
 
 /* Rudy Arthur's thesis pp.137
@@ -165,184 +329,238 @@ repeat
   →AVK =VKHK +fKe†K † Extend to an M = K + P step factorization AVM = VMHM + fMeM
 until convergence
 */
-  void calc(std::vector<RealD>& eval,  std::vector<Field>& evec, const Field& src, int& Nconv)
+  void calc(std::vector<RealD>& eval, std::vector<Field>& evec,  const Field& src, int& Nconv, bool reverse=false)
   {
+    GridBase *grid = src._grid;
+    assert(grid == evec[0]._grid);
     
-    GridBase *grid = evec[0]._grid;
+    GridLogIRL.TimingMode(1);
-    assert(grid == src._grid);
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
-    
+    std::cout << GridLogIRL <<" ImplicitlyRestartedLanczos::calc() starting iteration 0 /  "<< MaxIter<< std::endl;
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogMessage <<" ImplicitlyRestartedLanczos::calc() starting iteration 0 /  "<< MaxIter<< std::endl;
+    std::cout << GridLogIRL <<" -- seek   Nk    = " << Nk    <<" vectors"<< std::endl;
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL <<" -- accept Nstop = " << Nstop <<" vectors"<< std::endl;
-    std::cout << GridLogMessage <<" -- seek   Nk    = " << Nk    <<" vectors"<< std::endl;
+    std::cout << GridLogIRL <<" -- total  Nm    = " << Nm    <<" vectors"<< std::endl;
-    std::cout << GridLogMessage <<" -- accept Nstop = " << Nstop <<" vectors"<< std::endl;
+    std::cout << GridLogIRL <<" -- size of eval = " << eval.size() << std::endl;
-    std::cout << GridLogMessage <<" -- total  Nm    = " << Nm    <<" vectors"<< std::endl;
+    std::cout << GridLogIRL <<" -- size of evec = " << evec.size() << std::endl;
-    std::cout << GridLogMessage <<" -- size of eval = " << eval.size() << std::endl;
-    std::cout << GridLogMessage <<" -- size of evec = " << evec.size() << std::endl;
     if ( diagonalisation == IRLdiagonaliseWithDSTEGR ) {
-      std::cout << GridLogMessage << "Diagonalisation is DSTEGR "<<std::endl;
+      std::cout << GridLogIRL << "Diagonalisation is DSTEGR "<<std::endl;
     } else if ( diagonalisation == IRLdiagonaliseWithQR ) { 
-      std::cout << GridLogMessage << "Diagonalisation is QR "<<std::endl;
+      std::cout << GridLogIRL << "Diagonalisation is QR "<<std::endl;
     }  else if ( diagonalisation == IRLdiagonaliseWithEigen ) { 
-      std::cout << GridLogMessage << "Diagonalisation is Eigen "<<std::endl;
+      std::cout << GridLogIRL << "Diagonalisation is Eigen "<<std::endl;
     }
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
+	
+    assert(Nm <= evec.size() && Nm <= eval.size());
     
-    assert(Nm == evec.size() && Nm == eval.size());
+    // quickly get an idea of the largest eigenvalue to more properly normalize the residuum
+    RealD evalMaxApprox = 0.0;
+    {
+      auto src_n = src;
+      auto tmp = src;
+      const int _MAX_ITER_IRL_MEVAPP_ = 50;
+      for (int i=0;i<_MAX_ITER_IRL_MEVAPP_;i++) {
+	normalise(src_n);
+	_HermOp(src_n,tmp);
+	RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
+	RealD vden = norm2(src_n);
+	RealD na = vnum/vden;
+	if (fabs(evalMaxApprox/na - 1.0) < 0.05)
+	  i=_MAX_ITER_IRL_MEVAPP_;
+	evalMaxApprox = na;
+	std::cout << GridLogIRL << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
+	src_n = tmp;
+      }
+    }
 	
     std::vector<RealD> lme(Nm);  
     std::vector<RealD> lme2(Nm);
     std::vector<RealD> eval2(Nm);
+    std::vector<RealD> eval2_copy(Nm);
+    Eigen::MatrixXd Qt = Eigen::MatrixXd::Zero(Nm,Nm);
 
-    Eigen::MatrixXd    Qt = Eigen::MatrixXd::Zero(Nm,Nm);
-
-    std::vector<int>   Iconv(Nm);
-    std::vector<Field>  B(Nm,grid); // waste of space replicating
-    
     Field f(grid);
     Field v(grid);
-    
     int k1 = 1;
     int k2 = Nk;
-    
-    Nconv = 0;
-    
     RealD beta_k;
+
+    Nconv = 0;
   
     // Set initial vector
     evec[0] = src;
-    std::cout << GridLogMessage <<"norm2(src)= " << norm2(src)<<std::endl;
-    
     normalise(evec[0]);
-    std::cout << GridLogMessage <<"norm2(evec[0])= " << norm2(evec[0]) <<std::endl;
+	
-    
     // Initial Nk steps
+    OrthoTime=0.;
     for(int k=0; k<Nk; ++k) step(eval,lme,evec,f,Nm,k);
-    
+    std::cout<<GridLogIRL <<"Initial "<< Nk <<"steps done "<<std::endl;
+    std::cout<<GridLogIRL <<"Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
+
+    //////////////////////////////////
     // Restarting loop begins
+    //////////////////////////////////
     int iter;
     for(iter = 0; iter<MaxIter; ++iter){
       
+      OrthoTime=0.;
+
       std::cout<< GridLogMessage <<" **********************"<< std::endl;
       std::cout<< GridLogMessage <<" Restart iteration = "<< iter << std::endl;
       std::cout<< GridLogMessage <<" **********************"<< std::endl;
-      
+
+      std::cout<<GridLogIRL <<" running "<<Nm-Nk <<" steps: "<<std::endl;
       for(int k=Nk; k<Nm; ++k) step(eval,lme,evec,f,Nm,k);
-      
       f *= lme[Nm-1];
-      
+
+      std::cout<<GridLogIRL <<" "<<Nm-Nk <<" steps done "<<std::endl;
+      std::cout<<GridLogIRL <<"Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
+	  
+      //////////////////////////////////
       // getting eigenvalues
+      //////////////////////////////////
       for(int k=0; k<Nm; ++k){
 	eval2[k] = eval[k+k1-1];
 	lme2[k] = lme[k+k1-1];
       }
       Qt = Eigen::MatrixXd::Identity(Nm,Nm);
       diagonalize(eval2,lme2,Nm,Nm,Qt,grid);
+      std::cout<<GridLogIRL <<" diagonalized "<<std::endl;
 
+      //////////////////////////////////
       // sorting
-      _sort.push(eval2,Nm);
+      //////////////////////////////////
-      
+      eval2_copy = eval2;
+      std::partial_sort(eval2.begin(),eval2.begin()+Nm,eval2.end(),std::greater<RealD>());
+      std::cout<<GridLogIRL <<" evals sorted "<<std::endl;
+      const int chunk=8;
+      for(int io=0; io<k2;io+=chunk){
+	std::cout<<GridLogIRL << "eval "<< std::setw(3) << io ;
+	for(int ii=0;ii<chunk;ii++){
+	  if ( (io+ii)<k2 )
+	    std::cout<< " "<< std::setw(12)<< eval2[io+ii];
+	}
+	std::cout << std::endl;
+      }
+
+      //////////////////////////////////
       // Implicitly shifted QR transformations
+      //////////////////////////////////
       Qt = Eigen::MatrixXd::Identity(Nm,Nm);
       for(int ip=k2; ip<Nm; ++ip){ 
-	// Eigen replacement for qr_decomp ???
+	QR_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
-	qr_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
       }
-    
+      std::cout<<GridLogIRL <<"QR decomposed "<<std::endl;
-      for(int i=0; i<(Nk+1); ++i) B[i] = 0.0;
+
-	  
+      assert(k2<Nm);      assert(k2<Nm);      assert(k1>0);
-      for(int j=k1-1; j<k2+1; ++j){
+
-	for(int k=0; k<Nm; ++k){
+      basisRotate(evec,Qt,k1-1,k2+1,0,Nm,Nm); /// big constraint on the basis
-	  B[j].checkerboard = evec[k].checkerboard;
+      std::cout<<GridLogIRL <<"basisRotated  by Qt"<<std::endl;
-	  B[j] += Qt(j,k) * evec[k];
-	}
-      }
-      for(int j=k1-1; j<k2+1; ++j) evec[j] = B[j];
       
+      ////////////////////////////////////////////////////
       // Compressed vector f and beta(k2)
+      ////////////////////////////////////////////////////
       f *= Qt(k2-1,Nm-1);
       f += lme[k2-1] * evec[k2];
       beta_k = norm2(f);
       beta_k = sqrt(beta_k);
-      std::cout<< GridLogMessage<<" beta(k) = "<<beta_k<<std::endl;
+      std::cout<<GridLogIRL<<" beta(k) = "<<beta_k<<std::endl;
-      
+	  
       RealD betar = 1.0/beta_k;
       evec[k2] = betar * f;
       lme[k2-1] = beta_k;
-      
+	  
+      ////////////////////////////////////////////////////
       // Convergence test
+      ////////////////////////////////////////////////////
       for(int k=0; k<Nm; ++k){    
 	eval2[k] = eval[k];
 	lme2[k] = lme[k];
       }
       Qt = Eigen::MatrixXd::Identity(Nm,Nm);
       diagonalize(eval2,lme2,Nk,Nm,Qt,grid);
-      
+      std::cout<<GridLogIRL <<" Diagonalized "<<std::endl;
-      for(int k = 0; k<Nk; ++k) B[k]=0.0;
+	  
-      
-      for(int j = 0; j<Nk; ++j){
-	for(int k = 0; k<Nk; ++k){
-	  B[j].checkerboard = evec[k].checkerboard;
-	  B[j] += Qt(j,k) * evec[k];
-	}
-      }
-
       Nconv = 0;
-      for(int i=0; i<Nk; ++i){
+      if (iter >= MinRestart) {
-	
-	_Linop.HermOp(B[i],v);
-	    
-	RealD vnum = real(innerProduct(B[i],v)); // HermOp.
-	RealD vden = norm2(B[i]);
-	eval2[i] = vnum/vden;
-	v -= eval2[i]*B[i];
-	RealD vv = norm2(v);
-	
-	std::cout.precision(13);
-	std::cout << GridLogMessage << "[" << std::setw(3)<< std::setiosflags(std::ios_base::right) <<i<<"] ";
-	std::cout << "eval = "<<std::setw(25)<< std::setiosflags(std::ios_base::left)<< eval2[i];
-	std::cout << " |H B[i] - eval[i]B[i]|^2 "<< std::setw(25)<< std::setiosflags(std::ios_base::right)<< vv<< std::endl;
-	
-	// change the criteria as evals are supposed to be sorted, all evals smaller(larger) than Nstop should have converged
-	if((vv<eresid*eresid) && (i == Nconv) ){
-	  Iconv[Nconv] = i;
-	  ++Nconv;
-	}
-	
-      }  // i-loop end
-      
-      std::cout<< GridLogMessage <<" #modes converged: "<<Nconv<<std::endl;
 
-      if( Nconv>=Nstop ){
+	std::cout << GridLogIRL << "Test convergence: rotate subset of vectors to test convergence " << std::endl;
-	goto converged;
+
-      }
+	Field B(grid); B.checkerboard = evec[0].checkerboard;
-    } // end of iter loop
+
-    
+	//  power of two search pattern;  not every evalue in eval2 is assessed.
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
+	for(int jj = 1; jj<=Nstop; jj*=2){
-    std::cout<< GridLogError    <<" ImplicitlyRestartedLanczos::calc() NOT converged.";
+	  int j = Nstop-jj;
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
+	  RealD e = eval2_copy[j]; // Discard the evalue
+	  basisRotateJ(B,evec,Qt,j,0,Nk,Nm);	    
+	  if( _Tester.TestConvergence(j,eresid,B,e,evalMaxApprox) ) {
+	    if ( j > Nconv ) {
+	      Nconv=j+1;
+	      jj=Nstop; // Terminate the scan
+	    }
+	  }
+	}
+	// Do evec[0] for good measure
+	{ 
+	  int j=0;
+	  RealD e = eval2_copy[0]; 
+	  basisRotateJ(B,evec,Qt,j,0,Nk,Nm);	    
+	  _Tester.TestConvergence(j,eresid,B,e,evalMaxApprox);
+	}
+	// test if we converged, if so, terminate
+	std::cout<<GridLogIRL<<" #modes converged: >= "<<Nconv<<"/"<<Nstop<<std::endl;
+	//	if( Nconv>=Nstop || beta_k < betastp){
+	if( Nconv>=Nstop){
+	  goto converged;
+	}
+	  
+      } else {
+	std::cout << GridLogIRL << "iter < MinRestart: do not yet test for convergence\n";
+      } // end of iter loop
+    }
+
+    std::cout<<GridLogError<<"\n NOT converged.\n";
     abort();
 	
   converged:
-    // Sorting
+    {
-    eval.resize(Nconv);
+      Field B(grid); B.checkerboard = evec[0].checkerboard;
-    evec.resize(Nconv,grid);
+      basisRotate(evec,Qt,0,Nk,0,Nk,Nm);	    
-    for(int i=0; i<Nconv; ++i){
+      std::cout << GridLogIRL << " Rotated basis"<<std::endl;
-      eval[i] = eval2[Iconv[i]];
+      Nconv=0;
-      evec[i] = B[Iconv[i]];
+      //////////////////////////////////////////////////////////////////////
+      // Full final convergence test; unconditionally applied
+      //////////////////////////////////////////////////////////////////////
+      for(int j = 0; j<=Nk; j++){
+	B=evec[j];
+	if( _Tester.ReconstructEval(j,eresid,B,eval2[j],evalMaxApprox) ) {
+	  Nconv++;
+	}
+      }
+
+      if ( Nconv < Nstop )
+	std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
+
+      eval=eval2;
+      
+      //Keep only converged
+      eval.resize(Nconv);// Nstop?
+      evec.resize(Nconv,grid);// Nstop?
+      basisSortInPlace(evec,eval,reverse);
+      
     }
-    _sort.push(eval,evec,Nconv);
+       
-    
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL << "ImplicitlyRestartedLanczos CONVERGED ; Summary :\n";
-    std::cout << GridLogMessage << "ImplicitlyRestartedLanczos CONVERGED ; Summary :\n";
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
+    std::cout << GridLogIRL << " -- Iterations  = "<< iter   << "\n";
-    std::cout << GridLogMessage << " -- Iterations  = "<< iter   << "\n";
+    std::cout << GridLogIRL << " -- beta(k)     = "<< beta_k << "\n";
-    std::cout << GridLogMessage << " -- beta(k)     = "<< beta_k << "\n";
+    std::cout << GridLogIRL << " -- Nconv       = "<< Nconv  << "\n";
-    std::cout << GridLogMessage << " -- Nconv       = "<< Nconv  << "\n";
+    std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
-    std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
   }
 
-private:
+ private:
 /* Saad PP. 195
 1. Choose an initial vector v1 of 2-norm unity. Set β1 ≡ 0, v0 ≡ 0
 2. For k = 1,2,...,m Do:
@@ -360,28 +578,38 @@ private:
   {
     const RealD tiny = 1.0e-20;
     assert( k< Nm );
-    
+
-    _poly(_Linop,evec[k],w);      // 3. wk:=Avk−βkv_{k−1}
+    GridStopWatch gsw_op,gsw_o;
-    
+
+    Field& evec_k = evec[k];
+
+    _PolyOp(evec_k,w);    std::cout<<GridLogIRL << "PolyOp" <<std::endl;
+
     if(k>0) w -= lme[k-1] * evec[k-1];
-    
+
-    ComplexD zalph = innerProduct(evec[k],w); // 4. αk:=(wk,vk)
+    ComplexD zalph = innerProduct(evec_k,w); // 4. αk:=(wk,vk)
     RealD     alph = real(zalph);
-    
+
-    w = w - alph * evec[k];// 5. wk:=wk−αkvk
+    w = w - alph * evec_k;// 5. wk:=wk−αkvk
-    
+
     RealD beta = normalise(w); // 6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
     // 7. vk+1 := wk/βk+1
-    
+
     lmd[k] = alph;
     lme[k] = beta;
-    
+
-    if ( k > 0 ) orthogonalize(w,evec,k); // orthonormalise
+    if (k>0 && k % orth_period == 0) {
-    if ( k < Nm-1) evec[k+1] = w;
+      orthogonalize(w,evec,k); // orthonormalise
-    
+      std::cout<<GridLogIRL << "Orthogonalised " <<std::endl;
-    if ( beta < tiny ) std::cout << GridLogMessage << " beta is tiny "<<beta<<std::endl;
+    }
+
+    if(k < Nm-1) evec[k+1] = w;
+
+    std::cout<<GridLogIRL << "alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
+    if ( beta < tiny ) 
+      std::cout<<GridLogIRL << " beta is tiny "<<beta<<std::endl;
   }
-      
+
   void diagonalize_Eigen(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
 			 int Nk, int Nm,  
 			 Eigen::MatrixXd & Qt, // Nm x Nm
@@ -404,11 +632,11 @@ private:
       }
     }
   }
-  ///////////////////////////////////////////////////////////////////////////
-  // File could end here if settle on Eigen ???
-  ///////////////////////////////////////////////////////////////////////////
 
-  void qr_decomp(std::vector<RealD>& lmd,   // Nm 
+  ///////////////////////////////////////////////////////////////////////////
+  // File could end here if settle on Eigen ??? !!!
+  ///////////////////////////////////////////////////////////////////////////
+  void QR_decomp(std::vector<RealD>& lmd,   // Nm 
 		 std::vector<RealD>& lme,   // Nm 
 		 int Nk, int Nm,            // Nk, Nm
 		 Eigen::MatrixXd& Qt,       // Nm x Nm matrix
@@ -575,51 +803,50 @@ void diagonalize_lapack(std::vector<RealD>& lmd,
 #endif
 }
 
-  void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
+void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme, 
-		      int Nk, int Nm,   
+		    int Nk, int Nm,   
-		      Eigen::MatrixXd & Qt,
+		    Eigen::MatrixXd & Qt,
-		      GridBase *grid)
+		    GridBase *grid)
-  {
+{
-    int Niter = 100*Nm;
+  int QRiter = 100*Nm;
-    int kmin = 1;
+  int kmin = 1;
-    int kmax = Nk;
+  int kmax = Nk;
-
+  
-    // (this should be more sophisticated)
+  // (this should be more sophisticated)
-    for(int iter=0; iter<Niter; ++iter){
+  for(int iter=0; iter<QRiter; ++iter){
-      
+    
-      // determination of 2x2 leading submatrix
+    // determination of 2x2 leading submatrix
-      RealD dsub = lmd[kmax-1]-lmd[kmax-2];
+    RealD dsub = lmd[kmax-1]-lmd[kmax-2];
-      RealD dd = sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
+    RealD dd = sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
-      RealD Dsh = 0.5*(lmd[kmax-2]+lmd[kmax-1] +dd*(dsub/fabs(dsub)));
+    RealD Dsh = 0.5*(lmd[kmax-2]+lmd[kmax-1] +dd*(dsub/fabs(dsub)));
-      // (Dsh: shift)
+    // (Dsh: shift)
-	
+    
-      // transformation
+    // transformation
-      qr_decomp(lmd,lme,Nk,Nm,Qt,Dsh,kmin,kmax); // Nk, Nm
+    QR_decomp(lmd,lme,Nk,Nm,Qt,Dsh,kmin,kmax); // Nk, Nm
-	
+    
-      // Convergence criterion (redef of kmin and kamx)
+    // Convergence criterion (redef of kmin and kamx)
-      for(int j=kmax-1; j>= kmin; --j){
+    for(int j=kmax-1; j>= kmin; --j){
-	RealD dds = fabs(lmd[j-1])+fabs(lmd[j]);
+      RealD dds = fabs(lmd[j-1])+fabs(lmd[j]);
-	if(fabs(lme[j-1])+dds > dds){
+      if(fabs(lme[j-1])+dds > dds){
-	  kmax = j+1;
+	kmax = j+1;
-	  goto continued;
+	goto continued;
-	}
+      }
-      }
+    }
-      Niter = iter;
+    QRiter = iter;
-      return;
+    return;
-
+    
-    continued:
+  continued:
-      for(int j=0; j<kmax-1; ++j){
+    for(int j=0; j<kmax-1; ++j){
-	RealD dds = fabs(lmd[j])+fabs(lmd[j+1]);
+      RealD dds = fabs(lmd[j])+fabs(lmd[j+1]);
-	if(fabs(lme[j])+dds > dds){
+      if(fabs(lme[j])+dds > dds){
-	  kmin = j+1;
+	kmin = j+1;
-	  break;
+	break;
-	}
       }
     }
-    std::cout << GridLogError << "[QL method] Error - Too many iteration: "<<Niter<<"\n";
-    abort();
   }
-
+  std::cout << GridLogError << "[QL method] Error - Too many iteration: "<<QRiter<<"\n";
- };
+  abort();
+}
+};
 }
 #endif

lib/algorithms/iterative/LocalCoherenceLanczos.h

@@ -0,0 +1,352 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/LocalCoherenceLanczos.h
+
+    Copyright (C) 2015
+
+Author: Christoph Lehner <clehner@bnl.gov>
+Author: paboyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#ifndef GRID_LOCAL_COHERENCE_IRL_H
+#define GRID_LOCAL_COHERENCE_IRL_H
+namespace Grid { 
+struct LanczosParams : Serializable {
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParams,
+				  ChebyParams, Cheby,/*Chebyshev*/
+				  int, Nstop,    /*Vecs in Lanczos must converge Nstop < Nk < Nm*/
+				  int, Nk,       /*Vecs in Lanczos seek converge*/
+				  int, Nm,       /*Total vecs in Lanczos include restart*/
+				  RealD, resid,  /*residual*/
+ 				  int, MaxIt, 
+				  RealD, betastp,  /* ? */
+				  int, MinRes);    // Must restart
+};
+
+struct LocalCoherenceLanczosParams : Serializable {
+ public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
+				  bool, doFine,
+				  bool, doFineRead,
+				  bool, doCoarse,
+	       			  bool, doCoarseRead,
+				  LanczosParams, FineParams,
+				  LanczosParams, CoarseParams,
+				  ChebyParams,   Smoother,
+				  RealD        , coarse_relax_tol,
+				  std::vector<int>, blockSize,
+				  std::string, config,
+				  std::vector < std::complex<double>  >, omega,
+				  RealD, mass,
+				  RealD, M5);
+};
+
+// Duplicate functionality; ProjectedFunctionHermOp could be used with the trivial function
+template<class Fobj,class CComplex,int nbasis>
+class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
+public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj>          FineField;
+
+  LinearOperatorBase<FineField> &_Linop;
+  Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
+
+  ProjectedHermOp(LinearOperatorBase<FineField>& linop,  Aggregation<Fobj,CComplex,nbasis> &aggregate) : 
+    _Linop(linop),
+    _Aggregate(aggregate)  {  };
+
+  void operator()(const CoarseField& in, CoarseField& out) {
+
+    GridBase *FineGrid = _Aggregate.FineGrid;
+    FineField fin(FineGrid);
+    FineField fout(FineGrid);
+
+    _Aggregate.PromoteFromSubspace(in,fin);    std::cout<<GridLogIRL<<"ProjectedHermop : Promote to fine"<<std::endl;
+    _Linop.HermOp(fin,fout);                   std::cout<<GridLogIRL<<"ProjectedHermop : HermOp (fine) "<<std::endl;
+    _Aggregate.ProjectToSubspace(out,fout);    std::cout<<GridLogIRL<<"ProjectedHermop : Project to coarse "<<std::endl;
+  }
+};
+
+template<class Fobj,class CComplex,int nbasis>
+class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
+public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj>          FineField;
+
+
+  OperatorFunction<FineField>   & _poly;
+  LinearOperatorBase<FineField> &_Linop;
+  Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
+
+  ProjectedFunctionHermOp(OperatorFunction<FineField> & poly,LinearOperatorBase<FineField>& linop, 
+			  Aggregation<Fobj,CComplex,nbasis> &aggregate) : 
+    _poly(poly),
+    _Linop(linop),
+    _Aggregate(aggregate)  {  };
+
+  void operator()(const CoarseField& in, CoarseField& out) {
+
+    GridBase *FineGrid = _Aggregate.FineGrid;
+
+    FineField fin(FineGrid) ;fin.checkerboard  =_Aggregate.checkerboard;
+    FineField fout(FineGrid);fout.checkerboard =_Aggregate.checkerboard;
+    
+    _Aggregate.PromoteFromSubspace(in,fin);    std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Promote to fine"<<std::endl;
+    _poly(_Linop,fin,fout);                    std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Poly "<<std::endl;
+    _Aggregate.ProjectToSubspace(out,fout);    std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Project to coarse "<<std::endl;
+  }
+};
+
+template<class Fobj,class CComplex,int nbasis>
+class ImplicitlyRestartedLanczosSmoothedTester  : public ImplicitlyRestartedLanczosTester<Lattice<iVector<CComplex,nbasis > > >
+{
+ public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<CComplex>   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<Fobj>          FineField;
+
+  LinearFunction<CoarseField> & _Poly;
+  OperatorFunction<FineField>   & _smoother;
+  LinearOperatorBase<FineField> &_Linop;
+  Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
+  RealD                             _coarse_relax_tol;
+  ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField>   &Poly,
+					   OperatorFunction<FineField>   &smoother,
+					   LinearOperatorBase<FineField> &Linop,
+					   Aggregation<Fobj,CComplex,nbasis> &Aggregate,
+					   RealD coarse_relax_tol=5.0e3) 
+    : _smoother(smoother), _Linop(Linop),_Aggregate(Aggregate), _Poly(Poly), _coarse_relax_tol(coarse_relax_tol)  {    };
+
+  int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
+  {
+    CoarseField v(B);
+    RealD eval_poly = eval;
+    // Apply operator
+    _Poly(B,v);
+
+    RealD vnum = real(innerProduct(B,v)); // HermOp.
+    RealD vden = norm2(B);
+    RealD vv0  = norm2(v);
+    eval   = vnum/vden;
+    v -= eval*B;
+
+    RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
+
+    std::cout.precision(13);
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<std::endl;
+
+    int conv=0;
+    if( (vv<eresid*eresid) ) conv = 1;
+    return conv;
+  }
+  int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
+  {
+    GridBase *FineGrid = _Aggregate.FineGrid;
+
+    int checkerboard   = _Aggregate.checkerboard;
+
+    FineField fB(FineGrid);fB.checkerboard =checkerboard;
+    FineField fv(FineGrid);fv.checkerboard =checkerboard;
+
+    _Aggregate.PromoteFromSubspace(B,fv);
+    _smoother(_Linop,fv,fB); 
+
+    RealD eval_poly = eval;
+    _Linop.HermOp(fB,fv);
+
+    RealD vnum = real(innerProduct(fB,fv)); // HermOp.
+    RealD vden = norm2(fB);
+    RealD vv0  = norm2(fv);
+    eval   = vnum/vden;
+    fv -= eval*fB;
+    RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0);
+
+    std::cout.precision(13);
+    std::cout<<GridLogIRL  << "[" << std::setw(3)<<j<<"] "
+	     <<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
+	     <<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
+	     <<std::endl;
+    if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
+    if( (vv<eresid*eresid) ) return 1;
+    return 0;
+  }
+};
+
+////////////////////////////////////////////
+// Make serializable Lanczos params
+////////////////////////////////////////////
+template<class Fobj,class CComplex,int nbasis>
+class LocalCoherenceLanczos 
+{
+public:
+  typedef iVector<CComplex,nbasis >           CoarseSiteVector;
+  typedef Lattice<CComplex>                   CoarseScalar; // used for inner products on fine field
+  typedef Lattice<CoarseSiteVector>           CoarseField;
+  typedef Lattice<Fobj>                       FineField;
+
+protected:
+  GridBase *_CoarseGrid;
+  GridBase *_FineGrid;
+  int _checkerboard;
+  LinearOperatorBase<FineField>                 & _FineOp;
+  
+  // FIXME replace Aggregation with vector of fine; the code reuse is too small for
+  // the hassle and complexity of cross coupling.
+  Aggregation<Fobj,CComplex,nbasis>               _Aggregate;  
+  std::vector<RealD>                              evals_fine;
+  std::vector<RealD>                              evals_coarse; 
+  std::vector<CoarseField>                        evec_coarse;
+public:
+  LocalCoherenceLanczos(GridBase *FineGrid,
+		GridBase *CoarseGrid,
+		LinearOperatorBase<FineField> &FineOp,
+		int checkerboard) :
+    _CoarseGrid(CoarseGrid),
+    _FineGrid(FineGrid),
+    _Aggregate(CoarseGrid,FineGrid,checkerboard),
+    _FineOp(FineOp),
+    _checkerboard(checkerboard)
+  {
+    evals_fine.resize(0);
+    evals_coarse.resize(0);
+  };
+  void Orthogonalise(void ) { _Aggregate.Orthogonalise(); }
+
+  template<typename T>  static RealD normalise(T& v) 
+  {
+    RealD nn = norm2(v);
+    nn = ::sqrt(nn);
+    v = v * (1.0/nn);
+    return nn;
+  }
+
+  void fakeFine(void)
+  {
+    int Nk = nbasis;
+    _Aggregate.subspace.resize(Nk,_FineGrid);
+    _Aggregate.subspace[0]=1.0;
+    _Aggregate.subspace[0].checkerboard=_checkerboard;
+    normalise(_Aggregate.subspace[0]);
+    PlainHermOp<FineField>    Op(_FineOp);
+    for(int k=1;k<Nk;k++){
+      _Aggregate.subspace[k].checkerboard=_checkerboard;
+      Op(_Aggregate.subspace[k-1],_Aggregate.subspace[k]);
+      normalise(_Aggregate.subspace[k]);
+    }
+  }
+
+  void testFine(RealD resid) 
+  {
+    assert(evals_fine.size() == nbasis);
+    assert(_Aggregate.subspace.size() == nbasis);
+    PlainHermOp<FineField>    Op(_FineOp);
+    ImplicitlyRestartedLanczosHermOpTester<FineField> SimpleTester(Op);
+    for(int k=0;k<nbasis;k++){
+      assert(SimpleTester.ReconstructEval(k,resid,_Aggregate.subspace[k],evals_fine[k],1.0)==1);
+    }
+  }
+
+  void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax) 
+  {
+    assert(evals_fine.size() == nbasis);
+    assert(_Aggregate.subspace.size() == nbasis);
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+    // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+    Chebyshev<FineField>                          ChebySmooth(cheby_smooth);
+    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (ChebySmooth,_FineOp,_Aggregate);
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,_Aggregate,relax);
+
+    for(int k=0;k<evec_coarse.size();k++){
+      if ( k < nbasis ) { 
+	assert(ChebySmoothTester.ReconstructEval(k,resid,evec_coarse[k],evals_coarse[k],1.0)==1);
+      } else { 
+	assert(ChebySmoothTester.ReconstructEval(k,resid*relax,evec_coarse[k],evals_coarse[k],1.0)==1);
+      }
+    }
+  }
+
+  void calcFine(ChebyParams cheby_parms,int Nstop,int Nk,int Nm,RealD resid, 
+		RealD MaxIt, RealD betastp, int MinRes)
+  {
+    assert(nbasis<=Nm);
+    Chebyshev<FineField>      Cheby(cheby_parms);
+    FunctionHermOp<FineField> ChebyOp(Cheby,_FineOp);
+    PlainHermOp<FineField>    Op(_FineOp);
+
+    evals_fine.resize(Nm);
+    _Aggregate.subspace.resize(Nm,_FineGrid);
+
+    ImplicitlyRestartedLanczos<FineField> IRL(ChebyOp,Op,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
+
+    FineField src(_FineGrid); src=1.0; src.checkerboard = _checkerboard;
+
+    int Nconv;
+    IRL.calc(evals_fine,_Aggregate.subspace,src,Nconv,false);
+    
+    // Shrink down to number saved
+    assert(Nstop>=nbasis);
+    assert(Nconv>=nbasis);
+    evals_fine.resize(nbasis);
+    _Aggregate.subspace.resize(nbasis,_FineGrid);
+  }
+  void calcCoarse(ChebyParams cheby_op,ChebyParams cheby_smooth,RealD relax,
+		  int Nstop, int Nk, int Nm,RealD resid, 
+		  RealD MaxIt, RealD betastp, int MinRes)
+  {
+    Chebyshev<FineField>                          Cheby(cheby_op);
+    ProjectedHermOp<Fobj,CComplex,nbasis>         Op(_FineOp,_Aggregate);
+    ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,_Aggregate);
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+    // create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
+    //////////////////////////////////////////////////////////////////////////////////////////////////
+
+    Chebyshev<FineField>                                           ChebySmooth(cheby_smooth);
+    ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,_Aggregate,relax);
+
+    evals_coarse.resize(Nm);
+    evec_coarse.resize(Nm,_CoarseGrid);
+
+    CoarseField src(_CoarseGrid);     src=1.0; 
+
+    ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
+    int Nconv=0;
+    IRL.calc(evals_coarse,evec_coarse,src,Nconv,false);
+    assert(Nconv>=Nstop);
+    evals_coarse.resize(Nstop);
+    evec_coarse.resize (Nstop,_CoarseGrid);
+    for (int i=0;i<Nstop;i++){
+      std::cout << i << " Coarse eval = " << evals_coarse[i]  << std::endl;
+    }
+  }
+};
+
+}
+#endif

lib/algorithms/iterative/SchurRedBlack.h

@@ -53,16 +53,124 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
    *     M psi = eta
    ***********************
    *Odd
-   * i)   (D_oo)^{\dag} D_oo psi_o = (D_oo)^dag L^{-1}  eta_o
+   * i)                 D_oo psi_o =  L^{-1}  eta_o
    *                        eta_o' = (D_oo)^dag (eta_o - Moe Mee^{-1} eta_e)
+   *
+   * Wilson:
+   *      (D_oo)^{\dag} D_oo psi_o = (D_oo)^dag L^{-1}  eta_o
+   * Stag:
+   *      D_oo psi_o = L^{-1}  eta =    (eta_o - Moe Mee^{-1} eta_e)
+   *
+   * L^-1 eta_o= (1              0 ) (e
+   *             (-MoeMee^{-1}   1 )   
+   *
    *Even
    * ii)  Mee psi_e + Meo psi_o = src_e
    *
    *   => sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
    *
+   * 
+   * TODO: Other options:
+   * 
+   * a) change checkerboards for Schur e<->o
+   *
+   * Left precon by Moo^-1
+   * b) Doo^{dag} M_oo^-dag Moo^-1 Doo psi_0 =  (D_oo)^dag M_oo^-dag Moo^-1 L^{-1}  eta_o
+   *                              eta_o'     = (D_oo)^dag  M_oo^-dag Moo^-1 (eta_o - Moe Mee^{-1} eta_e)
+   *
+   * Right precon by Moo^-1
+   * c) M_oo^-dag Doo^{dag} Doo Moo^-1 phi_0 = M_oo^-dag (D_oo)^dag L^{-1}  eta_o
+   *                              eta_o'     = M_oo^-dag (D_oo)^dag (eta_o - Moe Mee^{-1} eta_e)
+   *                              psi_o = M_oo^-1 phi_o
+   * TODO: Deflation 
    */
 namespace Grid {
 
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Take a matrix and form a Red Black solver calling a Herm solver
+  // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Now make the norm reflect extra factor of Mee
+  template<class Field> class SchurRedBlackStaggeredSolve {
+  private:
+    OperatorFunction<Field> & _HermitianRBSolver;
+    int CBfactorise;
+  public:
+
+    /////////////////////////////////////////////////////
+    // Wrap the usual normal equations Schur trick
+    /////////////////////////////////////////////////////
+  SchurRedBlackStaggeredSolve(OperatorFunction<Field> &HermitianRBSolver)  :
+     _HermitianRBSolver(HermitianRBSolver) 
+    { 
+      CBfactorise=0;
+    };
+
+    template<class Matrix>
+      void operator() (Matrix & _Matrix,const Field &in, Field &out){
+
+      // FIXME CGdiagonalMee not implemented virtual function
+      // FIXME use CBfactorise to control schur decomp
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      SchurStaggeredOperator<Matrix,Field> _HermOpEO(_Matrix);
+ 
+      Field src_e(grid);
+      Field src_o(grid);
+      Field sol_e(grid);
+      Field sol_o(grid);
+      Field   tmp(grid);
+      Field  Mtmp(grid);
+      Field resid(fgrid);
+      
+      std::cout << GridLogMessage << " SchurRedBlackStaggeredSolve " <<std::endl;
+      pickCheckerboard(Even,src_e,in);
+      pickCheckerboard(Odd ,src_o,in);
+      pickCheckerboard(Even,sol_e,out);
+      pickCheckerboard(Odd ,sol_o,out);
+
+      std::cout << GridLogMessage << " SchurRedBlackStaggeredSolve checkerboards picked" <<std::endl;
+    
+      /////////////////////////////////////////////////////
+      // src_o = (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
+      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Odd);     
+
+      //src_o = tmp;     assert(src_o.checkerboard ==Odd);
+      _Matrix.Mooee(tmp,src_o); // Extra factor of "m" in source from dumb choice of matrix norm.
+
+      //////////////////////////////////////////////////////////////
+      // Call the red-black solver
+      //////////////////////////////////////////////////////////////
+      std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver calling the Mpc solver" <<std::endl;
+      _HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.checkerboard==Odd);
+      std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver called  the Mpc solver" <<std::endl;
+
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      ///////////////////////////////////////////////////
+      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Even);
+      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Even);
+      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Even);
+     
+      std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver reconstructed other CB" <<std::endl;
+      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Even);
+      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Odd );
+      std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver inserted solution" <<std::endl;
+
+      // Verify the unprec residual
+      _Matrix.M(out,resid); 
+      resid = resid-in;
+      RealD ns = norm2(in);
+      RealD nr = norm2(resid);
+      std::cout<<GridLogMessage << "SchurRedBlackStaggered solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
+    }     
+  };
+  template<class Field> using SchurRedBlackStagSolve = SchurRedBlackStaggeredSolve<Field>;
+
   ///////////////////////////////////////////////////////////////////////////////////////////////////////
   // Take a matrix and form a Red Black solver calling a Herm solver
   // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
@@ -76,12 +184,10 @@ namespace Grid {
     /////////////////////////////////////////////////////
     // Wrap the usual normal equations Schur trick
     /////////////////////////////////////////////////////
-  SchurRedBlackDiagMooeeSolve(OperatorFunction<Field> &HermitianRBSolver)  :
+  SchurRedBlackDiagMooeeSolve(OperatorFunction<Field> &HermitianRBSolver,int cb=0)  :  _HermitianRBSolver(HermitianRBSolver) 
-     _HermitianRBSolver(HermitianRBSolver) 
+  { 
-    { 
+    CBfactorise=cb;
-      CBfactorise=0;
+  };
-    };
-
     template<class Matrix>
       void operator() (Matrix & _Matrix,const Field &in, Field &out){
 
@@ -141,5 +247,166 @@ namespace Grid {
     }     
   };
 
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Take a matrix and form a Red Black solver calling a Herm solver
+  // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  template<class Field> class SchurRedBlackDiagTwoSolve {
+  private:
+    OperatorFunction<Field> & _HermitianRBSolver;
+    int CBfactorise;
+  public:
+
+    /////////////////////////////////////////////////////
+    // Wrap the usual normal equations Schur trick
+    /////////////////////////////////////////////////////
+  SchurRedBlackDiagTwoSolve(OperatorFunction<Field> &HermitianRBSolver)  :
+     _HermitianRBSolver(HermitianRBSolver) 
+    { 
+      CBfactorise=0;
+    };
+
+    template<class Matrix>
+      void operator() (Matrix & _Matrix,const Field &in, Field &out){
+
+      // FIXME CGdiagonalMee not implemented virtual function
+      // FIXME use CBfactorise to control schur decomp
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      SchurDiagTwoOperator<Matrix,Field> _HermOpEO(_Matrix);
+ 
+      Field src_e(grid);
+      Field src_o(grid);
+      Field sol_e(grid);
+      Field sol_o(grid);
+      Field   tmp(grid);
+      Field  Mtmp(grid);
+      Field resid(fgrid);
+
+      pickCheckerboard(Even,src_e,in);
+      pickCheckerboard(Odd ,src_o,in);
+      pickCheckerboard(Even,sol_e,out);
+      pickCheckerboard(Odd ,sol_o,out);
+    
+      /////////////////////////////////////////////////////
+      // src_o = Mdag * (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
+      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Odd);     
+
+      // get the right MpcDag
+      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.checkerboard ==Odd);       
+
+      //////////////////////////////////////////////////////////////
+      // Call the red-black solver
+      //////////////////////////////////////////////////////////////
+      std::cout<<GridLogMessage << "SchurRedBlack solver calling the MpcDagMp solver" <<std::endl;
+//      _HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.checkerboard==Odd);
+      _HermitianRBSolver(_HermOpEO,src_o,tmp);  assert(tmp.checkerboard==Odd);
+      _Matrix.MooeeInv(tmp,sol_o);        assert(  sol_o.checkerboard   ==Odd);
+
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      ///////////////////////////////////////////////////
+      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Even);
+      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Even);
+      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Even);
+     
+      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Even);
+      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Odd );
+
+      // Verify the unprec residual
+      _Matrix.M(out,resid); 
+      resid = resid-in;
+      RealD ns = norm2(in);
+      RealD nr = norm2(resid);
+
+      std::cout<<GridLogMessage << "SchurRedBlackDiagTwo solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
+    }     
+  };
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Take a matrix and form a Red Black solver calling a Herm solver
+  // Use of RB info prevents making SchurRedBlackSolve conform to standard interface
+  ///////////////////////////////////////////////////////////////////////////////////////////////////////
+  template<class Field> class SchurRedBlackDiagTwoMixed {
+  private:
+    LinearFunction<Field> & _HermitianRBSolver;
+    int CBfactorise;
+  public:
+
+    /////////////////////////////////////////////////////
+    // Wrap the usual normal equations Schur trick
+    /////////////////////////////////////////////////////
+  SchurRedBlackDiagTwoMixed(LinearFunction<Field> &HermitianRBSolver)  :
+     _HermitianRBSolver(HermitianRBSolver) 
+    { 
+      CBfactorise=0;
+    };
+
+    template<class Matrix>
+      void operator() (Matrix & _Matrix,const Field &in, Field &out){
+
+      // FIXME CGdiagonalMee not implemented virtual function
+      // FIXME use CBfactorise to control schur decomp
+      GridBase *grid = _Matrix.RedBlackGrid();
+      GridBase *fgrid= _Matrix.Grid();
+
+      SchurDiagTwoOperator<Matrix,Field> _HermOpEO(_Matrix);
+ 
+      Field src_e(grid);
+      Field src_o(grid);
+      Field sol_e(grid);
+      Field sol_o(grid);
+      Field   tmp(grid);
+      Field  Mtmp(grid);
+      Field resid(fgrid);
+
+      pickCheckerboard(Even,src_e,in);
+      pickCheckerboard(Odd ,src_o,in);
+      pickCheckerboard(Even,sol_e,out);
+      pickCheckerboard(Odd ,sol_o,out);
+    
+      /////////////////////////////////////////////////////
+      // src_o = Mdag * (source_o - Moe MeeInv source_e)
+      /////////////////////////////////////////////////////
+      _Matrix.MooeeInv(src_e,tmp);     assert(  tmp.checkerboard ==Even);
+      _Matrix.Meooe   (tmp,Mtmp);      assert( Mtmp.checkerboard ==Odd);     
+      tmp=src_o-Mtmp;                  assert(  tmp.checkerboard ==Odd);     
+
+      // get the right MpcDag
+      _HermOpEO.MpcDag(tmp,src_o);     assert(src_o.checkerboard ==Odd);       
+
+      //////////////////////////////////////////////////////////////
+      // Call the red-black solver
+      //////////////////////////////////////////////////////////////
+      std::cout<<GridLogMessage << "SchurRedBlack solver calling the MpcDagMp solver" <<std::endl;
+//      _HermitianRBSolver(_HermOpEO,src_o,sol_o);  assert(sol_o.checkerboard==Odd);
+//      _HermitianRBSolver(_HermOpEO,src_o,tmp);  assert(tmp.checkerboard==Odd);
+      _HermitianRBSolver(src_o,tmp);  assert(tmp.checkerboard==Odd);
+      _Matrix.MooeeInv(tmp,sol_o);        assert(  sol_o.checkerboard   ==Odd);
+
+      ///////////////////////////////////////////////////
+      // sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
+      ///////////////////////////////////////////////////
+      _Matrix.Meooe(sol_o,tmp);        assert(  tmp.checkerboard   ==Even);
+      src_e = src_e-tmp;               assert(  src_e.checkerboard ==Even);
+      _Matrix.MooeeInv(src_e,sol_e);   assert(  sol_e.checkerboard ==Even);
+     
+      setCheckerboard(out,sol_e); assert(  sol_e.checkerboard ==Even);
+      setCheckerboard(out,sol_o); assert(  sol_o.checkerboard ==Odd );
+
+      // Verify the unprec residual
+      _Matrix.M(out,resid); 
+      resid = resid-in;
+      RealD ns = norm2(in);
+      RealD nr = norm2(resid);
+
+      std::cout<<GridLogMessage << "SchurRedBlackDiagTwo solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
+    }     
+  };
+
 }
 #endif

lib/allocator/AlignedAllocator.cc

@@ -3,9 +3,12 @@
 
 namespace Grid {
 
+MemoryStats *MemoryProfiler::stats = nullptr;
+bool         MemoryProfiler::debug = false;
+
 int PointerCache::victim;
 
-  PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
+PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
 
 void *PointerCache::Insert(void *ptr,size_t bytes) {
 
@@ -94,4 +97,29 @@ void check_huge_pages(void *Buf,uint64_t BYTES)
 #endif
 }
 
+std::string sizeString(const size_t bytes)
+{
+  constexpr unsigned int bufSize = 256;
+  const char             *suffixes[7] = {"", "K", "M", "G", "T", "P", "E"};
+  char                   buf[256];
+  size_t                 s     = 0;
+  double                 count = bytes;
+  
+  while (count >= 1024 && s < 7)
+  {
+      s++;
+      count /= 1024;
+  }
+  if (count - floor(count) == 0.0)
+  {
+      snprintf(buf, bufSize, "%d %sB", (int)count, suffixes[s]);
+  }
+  else
+  {
+      snprintf(buf, bufSize, "%.1f %sB", count, suffixes[s]);
+  }
+  
+  return std::string(buf);
+}
+
 }

lib/allocator/AlignedAllocator.h

@@ -63,6 +63,64 @@ namespace Grid {
     static void *Lookup(size_t bytes) ;
 
   };
+  
+  std::string sizeString(size_t bytes);
+
+  struct MemoryStats
+  {
+    size_t totalAllocated{0}, maxAllocated{0}, 
+           currentlyAllocated{0}, totalFreed{0};
+  };
+    
+  class MemoryProfiler
+  {
+  public:
+    static MemoryStats *stats;
+    static bool        debug;
+  };
+
+  #define memString(bytes) std::to_string(bytes) + " (" + sizeString(bytes) + ")"
+  #define profilerDebugPrint \
+  if (MemoryProfiler::stats)\
+  {\
+    auto s = MemoryProfiler::stats;\
+    std::cout << GridLogDebug << "[Memory debug] Stats " << MemoryProfiler::stats << std::endl;\
+    std::cout << GridLogDebug << "[Memory debug] total  : " << memString(s->totalAllocated) \
+              << std::endl;\
+    std::cout << GridLogDebug << "[Memory debug] max    : " << memString(s->maxAllocated) \
+              << std::endl;\
+    std::cout << GridLogDebug << "[Memory debug] current: " << memString(s->currentlyAllocated) \
+              << std::endl;\
+    std::cout << GridLogDebug << "[Memory debug] freed  : " << memString(s->totalFreed) \
+              << std::endl;\
+  }
+
+  #define profilerAllocate(bytes)\
+  if (MemoryProfiler::stats)\
+  {\
+    auto s = MemoryProfiler::stats;\
+    s->totalAllocated     += (bytes);\
+    s->currentlyAllocated += (bytes);\
+    s->maxAllocated        = std::max(s->maxAllocated, s->currentlyAllocated);\
+  }\
+  if (MemoryProfiler::debug)\
+  {\
+    std::cout << GridLogDebug << "[Memory debug] allocating " << memString(bytes) << std::endl;\
+    profilerDebugPrint;\
+  }
+
+  #define profilerFree(bytes)\
+  if (MemoryProfiler::stats)\
+  {\
+    auto s = MemoryProfiler::stats;\
+    s->totalFreed         += (bytes);\
+    s->currentlyAllocated -= (bytes);\
+  }\
+  if (MemoryProfiler::debug)\
+  {\
+    std::cout << GridLogDebug << "[Memory debug] freeing " << memString(bytes) << std::endl;\
+    profilerDebugPrint;\
+  }
 
   void check_huge_pages(void *Buf,uint64_t BYTES);
 
@@ -92,6 +150,7 @@ public:
   pointer allocate(size_type __n, const void* _p= 0)
   { 
     size_type bytes = __n*sizeof(_Tp);
+    profilerAllocate(bytes);
 
     _Tp *ptr = (_Tp *) PointerCache::Lookup(bytes);
     //    if ( ptr != NULL ) 
@@ -122,6 +181,8 @@ public:
   void deallocate(pointer __p, size_type __n) { 
     size_type bytes = __n * sizeof(_Tp);
 
+    profilerFree(bytes);
+
     pointer __freeme = (pointer)PointerCache::Insert((void *)__p,bytes);
 
 #ifdef HAVE_MM_MALLOC_H
@@ -172,10 +233,13 @@ public:
 #ifdef GRID_COMMS_SHMEM
   pointer allocate(size_type __n, const void* _p= 0)
   {
+    size_type bytes = __n*sizeof(_Tp);
+
+    profilerAllocate(bytes);
 #ifdef CRAY
-    _Tp *ptr = (_Tp *) shmem_align(__n*sizeof(_Tp),64);
+    _Tp *ptr = (_Tp *) shmem_align(bytes,64);
 #else
-    _Tp *ptr = (_Tp *) shmem_align(64,__n*sizeof(_Tp));
+    _Tp *ptr = (_Tp *) shmem_align(64,bytes);
 #endif
 #ifdef PARANOID_SYMMETRIC_HEAP
     static void * bcast;
@@ -193,18 +257,23 @@ public:
 #endif 
     return ptr;
   }
-  void deallocate(pointer __p, size_type) { 
+  void deallocate(pointer __p, size_type __n) { 
+    size_type bytes = __n*sizeof(_Tp);
+
+    profilerFree(bytes);
     shmem_free((void *)__p);
   }
 #else
   pointer allocate(size_type __n, const void* _p= 0) 
   {
-#ifdef HAVE_MM_MALLOC_H
-    _Tp * ptr = (_Tp *) _mm_malloc(__n*sizeof(_Tp),GRID_ALLOC_ALIGN);
-#else
-    _Tp * ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN,__n*sizeof(_Tp));
-#endif
     size_type bytes = __n*sizeof(_Tp);
+    
+    profilerAllocate(bytes);
+#ifdef HAVE_MM_MALLOC_H
+    _Tp * ptr = (_Tp *) _mm_malloc(bytes, GRID_ALLOC_ALIGN);
+#else
+    _Tp * ptr = (_Tp *) memalign(GRID_ALLOC_ALIGN, bytes);
+#endif
     uint8_t *cp = (uint8_t *)ptr;
     if ( ptr ) { 
     // One touch per 4k page, static OMP loop to catch same loop order
@@ -215,7 +284,10 @@ public:
     }
     return ptr;
   }
-  void deallocate(pointer __p, size_type) { 
+  void deallocate(pointer __p, size_type __n) {
+    size_type bytes = __n*sizeof(_Tp);
+
+    profilerFree(bytes);
 #ifdef HAVE_MM_MALLOC_H
     _mm_free((void *)__p); 
 #else

lib/cartesian/Cartesian_base.h

@@ -44,13 +44,21 @@ namespace Grid{
   class GridBase : public CartesianCommunicator , public GridThread {
 
 public:
-
+    int dummy;
     // Give Lattice access
     template<class object> friend class Lattice;
 
     GridBase(const std::vector<int> & processor_grid) : CartesianCommunicator(processor_grid) {};
     GridBase(const std::vector<int> & processor_grid,
-	     const CartesianCommunicator &parent) : CartesianCommunicator(processor_grid,parent) {};
+	     const CartesianCommunicator &parent,
+	     int &split_rank) 
+      : CartesianCommunicator(processor_grid,parent,split_rank) {};
+    GridBase(const std::vector<int> & processor_grid,
+	     const CartesianCommunicator &parent) 
+      : CartesianCommunicator(processor_grid,parent,dummy) {};
+
+    virtual ~GridBase() = default;
+
 
     // Physics Grid information.
     std::vector<int> _simd_layout;// Which dimensions get relayed out over simd lanes.

lib/cartesian/Cartesian_full.h

@@ -38,7 +38,7 @@ namespace Grid{
 class GridCartesian: public GridBase {
 
 public:
-
+    int dummy;
     virtual int  CheckerBoardFromOindexTable (int Oindex) {
       return 0;
     }
@@ -67,7 +67,14 @@ public:
     GridCartesian(const std::vector<int> &dimensions,
 		  const std::vector<int> &simd_layout,
 		  const std::vector<int> &processor_grid,
-		  const GridCartesian &parent) : GridBase(processor_grid,parent)
+		  const GridCartesian &parent) : GridBase(processor_grid,parent,dummy)
+    {
+      Init(dimensions,simd_layout,processor_grid);
+    }
+    GridCartesian(const std::vector<int> &dimensions,
+		  const std::vector<int> &simd_layout,
+		  const std::vector<int> &processor_grid,
+		  const GridCartesian &parent,int &split_rank) : GridBase(processor_grid,parent,split_rank)
     {
       Init(dimensions,simd_layout,processor_grid);
     }
@@ -81,6 +88,8 @@ public:
       Init(dimensions,simd_layout,processor_grid);
     }
 
+    virtual ~GridCartesian() = default;
+
     void Init(const std::vector<int> &dimensions,
 	      const std::vector<int> &simd_layout,
 	      const std::vector<int> &processor_grid)
@@ -113,6 +122,7 @@ public:
 
         // Use a reduced simd grid
         _ldimensions[d] = _gdimensions[d] / _processors[d]; //local dimensions
+        //std::cout << _ldimensions[d] << "  " << _gdimensions[d] << "  " << _processors[d] << std::endl;
         assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);
 
         _rdimensions[d] = _ldimensions[d] / _simd_layout[d]; //overdecomposition
@@ -157,6 +167,7 @@ public:
         block = block * _rdimensions[d];
       }
     };
+
 };
 }
 #endif

lib/cartesian/Cartesian_red_black.h

@@ -133,6 +133,8 @@ public:
     {
       Init(base->_fdimensions,base->_simd_layout,base->_processors,checker_dim_mask,checker_dim)  ;
     }
+
+    virtual ~GridRedBlackCartesian() = default;
 #if 0
     ////////////////////////////////////////////////////////////
     // Create redblack grid ;; deprecate these. Should not
@@ -205,6 +207,7 @@ public:
         {
           assert((_gdimensions[d] & 0x1) == 0);
           _gdimensions[d] = _gdimensions[d] / 2; // Remove a checkerboard
+	  _gsites /= 2;
         }
         _ldimensions[d] = _gdimensions[d] / _processors[d];
         assert(_ldimensions[d] * _processors[d] == _gdimensions[d]);

lib/communicator/Communicator.h

@@ -28,6 +28,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_COMMUNICATOR_H
 #define GRID_COMMUNICATOR_H
 
+#include <Grid/communicator/SharedMemory.h>
 #include <Grid/communicator/Communicator_base.h>
 
 #endif

lib/communicator/Communicator_base.cc

@@ -36,33 +36,9 @@ namespace Grid {
 ///////////////////////////////////////////////////////////////
 // Info that is setup once and indept of cartesian layout
 ///////////////////////////////////////////////////////////////
-void *              CartesianCommunicator::ShmCommBuf;
-uint64_t            CartesianCommunicator::MAX_MPI_SHM_BYTES   = 1024LL*1024LL*1024LL; 
 CartesianCommunicator::CommunicatorPolicy_t  
 CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
 int CartesianCommunicator::nCommThreads = -1;
-int CartesianCommunicator::Hugepages = 0;
-
-/////////////////////////////////
-// Alloc, free shmem region
-/////////////////////////////////
-void *CartesianCommunicator::ShmBufferMalloc(size_t bytes){
-  //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
-  void *ptr = (void *)heap_top;
-  heap_top  += bytes;
-  heap_bytes+= bytes;
-  if (heap_bytes >= MAX_MPI_SHM_BYTES) {
-    std::cout<< " ShmBufferMalloc exceeded shared heap size -- try increasing with --shm <MB> flag" <<std::endl;
-    std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
-    std::cout<< " Current value is " << (MAX_MPI_SHM_BYTES/(1024*1024)) <<std::endl;
-    assert(heap_bytes<MAX_MPI_SHM_BYTES);
-  }
-  return ptr;
-}
-void CartesianCommunicator::ShmBufferFreeAll(void) { 
-  heap_top  =(size_t)ShmBufferSelf();
-  heap_bytes=0;
-}
 
 /////////////////////////////////
 // Grid information queries
@@ -95,183 +71,6 @@ void CartesianCommunicator::GlobalSumVector(ComplexD *c,int N)
 {
   GlobalSumVector((double *)c,2*N);
 }
-
-
-#if defined( GRID_COMMS_MPI) || defined (GRID_COMMS_MPIT)
-
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent) 
-{
-  _ndimension = processors.size();
-  assert(_ndimension = parent._ndimension);
-  
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  // split the communicator
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  int Nparent;
-  MPI_Comm_size(parent.communicator,&Nparent);
-
-  int childsize=1;
-  for(int d=0;d<processors.size();d++) {
-    childsize *= processors[d];
-  }
-  int Nchild = Nparent/childsize;
-  assert (childsize * Nchild == Nparent);
-
-  int prank;  MPI_Comm_rank(parent.communicator,&prank);
-  int crank = prank % childsize;
-  int ccomm = prank / childsize;
-
-  MPI_Comm comm_split;
-  if ( Nchild > 1 ) { 
-
-    std::cout << GridLogMessage<<"Child communicator of "<< std::hex << parent.communicator << std::dec<<std::endl;
-    std::cout << GridLogMessage<<" parent grid["<< parent._ndimension<<"]    ";
-    for(int d=0;d<parent._processors.size();d++)  std::cout << parent._processors[d] << " ";
-    std::cout<<std::endl;
-
-    std::cout << GridLogMessage<<" child grid["<< _ndimension <<"]    ";
-    for(int d=0;d<processors.size();d++)  std::cout << processors[d] << " ";
-    std::cout<<std::endl;
-
-    int ierr= MPI_Comm_split(parent.communicator, ccomm,crank,&comm_split);
-    assert(ierr==0);
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    // Declare victory
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    std::cout << GridLogMessage<<"Divided communicator "<< parent._Nprocessors<<" into "
-	      <<Nchild <<" communicators with " << childsize << " ranks"<<std::endl;
-  } else {
-    comm_split=parent.communicator;
-    //    std::cout << "Passed parental communicator to a new communicator" <<std::endl;
-  }
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Set up from the new split communicator
-  //////////////////////////////////////////////////////////////////////////////////////////////////////
-  InitFromMPICommunicator(processors,comm_split);
-}
-//////////////////////////////////////////////////////////////////////////////////////////////////////
-// Take an MPI_Comm and self assemble
-//////////////////////////////////////////////////////////////////////////////////////////////////////
-void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base)
-{
-  //  if ( communicator_base != communicator_world ) {
-  //    std::cout << "Cartesian communicator created with a non-world communicator"<<std::endl;
-  //  }
-  _ndimension = processors.size();
-  _processor_coor.resize(_ndimension);
-
-  /////////////////////////////////
-  // Count the requested nodes
-  /////////////////////////////////
-  _Nprocessors=1;
-  _processors = processors;
-  for(int i=0;i<_ndimension;i++){
-    _Nprocessors*=_processors[i];
-  }
-
-  std::vector<int> periodic(_ndimension,1);
-  MPI_Cart_create(communicator_base, _ndimension,&_processors[0],&periodic[0],1,&communicator);
-  MPI_Comm_rank(communicator,&_processor);
-  MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
-
-  int Size;
-  MPI_Comm_size(communicator,&Size);
-
-#ifdef GRID_COMMS_MPIT
-  communicator_halo.resize (2*_ndimension);
-  for(int i=0;i<_ndimension*2;i++){
-    MPI_Comm_dup(communicator,&communicator_halo[i]);
-  }
-#endif
-  
-  assert(Size==_Nprocessors);
-}
-
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors) 
-{
-  InitFromMPICommunicator(processors,communicator_world);
-}
-
-#endif
-
-#if !defined( GRID_COMMS_MPI3) 
-
-int                      CartesianCommunicator::NodeCount(void)    { return ProcessorCount();};
-int                      CartesianCommunicator::RankCount(void)    { return ProcessorCount();};
-#endif
-#if !defined( GRID_COMMS_MPI3) && !defined (GRID_COMMS_MPIT)
-double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
-						     int xmit_to_rank,
-						     void *recv,
-						     int recv_from_rank,
-						     int bytes, int dir)
-{
-  std::vector<CommsRequest_t> list;
-  // Discard the "dir"
-  SendToRecvFromBegin   (list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
-  SendToRecvFromComplete(list);
-  return 2.0*bytes;
-}
-double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-							 void *xmit,
-							 int xmit_to_rank,
-							 void *recv,
-							 int recv_from_rank,
-							 int bytes, int dir)
-{
-  // Discard the "dir"
-  SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
-  return 2.0*bytes;
-}
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
-{
-  SendToRecvFromComplete(waitall);
-}
-#endif
-
-#if !defined( GRID_COMMS_MPI3) 
-
-void CartesianCommunicator::StencilBarrier(void){};
-
-commVector<uint8_t> CartesianCommunicator::ShmBufStorageVector;
-
-void *CartesianCommunicator::ShmBufferSelf(void) { return ShmCommBuf; }
-
-void *CartesianCommunicator::ShmBuffer(int rank) {
-  return NULL;
-}
-void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p) { 
-  return NULL;
-}
-void CartesianCommunicator::ShmInitGeneric(void){
-#if 1
-  int mmap_flag =0;
-#ifdef MAP_ANONYMOUS
-  mmap_flag = mmap_flag| MAP_SHARED | MAP_ANONYMOUS;
-#endif
-#ifdef MAP_ANON
-  mmap_flag = mmap_flag| MAP_SHARED | MAP_ANON;
-#endif
-#ifdef MAP_HUGETLB
-  if ( Hugepages ) mmap_flag |= MAP_HUGETLB;
-#endif
-  ShmCommBuf =(void *) mmap(NULL, MAX_MPI_SHM_BYTES, PROT_READ | PROT_WRITE, mmap_flag, -1, 0); 
-  if (ShmCommBuf == (void *)MAP_FAILED) {
-    perror("mmap failed ");
-    exit(EXIT_FAILURE);  
-  }
-#ifdef MADV_HUGEPAGE
-  if (!Hugepages ) madvise(ShmCommBuf,MAX_MPI_SHM_BYTES,MADV_HUGEPAGE);
-#endif
-#else 
-  ShmBufStorageVector.resize(MAX_MPI_SHM_BYTES);
-  ShmCommBuf=(void *)&ShmBufStorageVector[0];
-#endif
-  bzero(ShmCommBuf,MAX_MPI_SHM_BYTES);
-}
-
-#endif
   
 }
 

lib/communicator/Communicator_base.h

@@ -32,117 +32,33 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 ///////////////////////////////////
 // Processor layout information
 ///////////////////////////////////
-#ifdef GRID_COMMS_MPI
+#include <Grid/communicator/SharedMemory.h>
-#include <mpi.h>
-#endif
-#ifdef GRID_COMMS_MPI3
-#include <mpi.h>
-#endif
-#ifdef GRID_COMMS_MPIT
-#include <mpi.h>
-#endif
-#ifdef GRID_COMMS_SHMEM
-#include <mpp/shmem.h>
-#endif
 
 namespace Grid {
 
-class CartesianCommunicator {
+class CartesianCommunicator : public SharedMemory {
-  public:    
 
+public:    
 
   ////////////////////////////////////////////
-  // Isend/Irecv/Wait, or Sendrecv blocking
+  // Policies
   ////////////////////////////////////////////
   enum CommunicatorPolicy_t { CommunicatorPolicyConcurrent, CommunicatorPolicySequential };
   static CommunicatorPolicy_t CommunicatorPolicy;
   static void SetCommunicatorPolicy(CommunicatorPolicy_t policy ) { CommunicatorPolicy = policy; }
-
-  ///////////////////////////////////////////
-  // Up to 65536 ranks per node adequate for now
-  // 128MB shared memory for comms enought for 48^4 local vol comms
-  // Give external control (command line override?) of this
-  ///////////////////////////////////////////
-  static const int MAXLOG2RANKSPERNODE = 16;            
-  static uint64_t  MAX_MPI_SHM_BYTES;
   static int       nCommThreads;
-  // use explicit huge pages
-  static int       Hugepages;
 
+  ////////////////////////////////////////////
   // Communicator should know nothing of the physics grid, only processor grid.
+  ////////////////////////////////////////////
   int              _Nprocessors;     // How many in all
   std::vector<int> _processors;      // Which dimensions get relayed out over processors lanes.
   int              _processor;       // linear processor rank
   std::vector<int> _processor_coor;  // linear processor coordinate
-  unsigned long _ndimension;
+  unsigned long    _ndimension;
-
+  static Grid_MPI_Comm      communicator_world;
-#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPI3) || defined (GRID_COMMS_MPIT)
+  Grid_MPI_Comm             communicator;
-  static MPI_Comm communicator_world;
+  std::vector<Grid_MPI_Comm> communicator_halo;
-
-  MPI_Comm              communicator;
-  std::vector<MPI_Comm> communicator_halo;
-
-  typedef MPI_Request CommsRequest_t;
-
-#else 
-  typedef int CommsRequest_t;
-#endif
-
-
-  ////////////////////////////////////////////////////////////////////
-  // Helper functionality for SHM Windows common to all other impls
-  ////////////////////////////////////////////////////////////////////
-  // Longer term; drop this in favour of a master / slave model with 
-  // cartesian communicator on a subset of ranks, slave ranks controlled
-  // by group leader with data xfer via shared memory
-  ////////////////////////////////////////////////////////////////////
-#ifdef GRID_COMMS_MPI3
-
-  static int ShmRank;
-  static int ShmSize;
-  static int GroupRank;
-  static int GroupSize;
-  static int WorldRank;
-  static int WorldSize;
-
-  std::vector<int>  WorldDims;
-  std::vector<int>  GroupDims;
-  std::vector<int>  ShmDims;
-  
-  std::vector<int> GroupCoor;
-  std::vector<int> ShmCoor;
-  std::vector<int> WorldCoor;
-
-  static std::vector<int> GroupRanks; 
-  static std::vector<int> MyGroup;
-  static int ShmSetup;
-  static MPI_Win ShmWindow; 
-  static MPI_Comm ShmComm;
-  
-  std::vector<int>  LexicographicToWorldRank;
-  
-  static std::vector<void *> ShmCommBufs;
-
-#else 
-  static void ShmInitGeneric(void);
-  static commVector<uint8_t> ShmBufStorageVector;
-#endif 
-
-  /////////////////////////////////
-  // Grid information and queries
-  // Implemented in Communicator_base.C
-  /////////////////////////////////
-  static void * ShmCommBuf;
-
-  
-  size_t heap_top;
-  size_t heap_bytes;
-
-  void *ShmBufferSelf(void);
-  void *ShmBuffer(int rank);
-  void *ShmBufferTranslate(int rank,void * local_p);
-  void *ShmBufferMalloc(size_t bytes);
-  void ShmBufferFreeAll(void) ;
   
   ////////////////////////////////////////////////
   // Must call in Grid startup
@@ -153,18 +69,20 @@ class CartesianCommunicator {
   // Constructors to sub-divide a parent communicator
   // and default to comm world
   ////////////////////////////////////////////////
-  CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent);
+  CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank);
   CartesianCommunicator(const std::vector<int> &pdimensions_in);
+  virtual ~CartesianCommunicator();
 
  private:
-#if defined (GRID_COMMS_MPI) || defined (GRID_COMMS_MPIT) 
+
   ////////////////////////////////////////////////
   // Private initialise from an MPI communicator
   // Can use after an MPI_Comm_split, but hidden from user so private
   ////////////////////////////////////////////////
-  void InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base);
+  void InitFromMPICommunicator(const std::vector<int> &processors, Grid_MPI_Comm communicator_base);
-#endif
+
  public:
+
   
   ////////////////////////////////////////////////////////////////////////////////////////
   // Wraps MPI_Cart routines, or implements equivalent on other impls
@@ -180,8 +98,6 @@ class CartesianCommunicator {
   const std::vector<int> & ThisProcessorCoor(void) ;
   const std::vector<int> & ProcessorGrid(void)     ;
   int                      ProcessorCount(void)    ;
-  int                      NodeCount(void)    ;
-  int                      RankCount(void)    ;
 
   ////////////////////////////////////////////////////////////////////////////////
   // very VERY rarely (Log, serial RNG) we need world without a grid
@@ -262,6 +178,23 @@ class CartesianCommunicator {
   // Broadcast a buffer and composite larger
   ////////////////////////////////////////////////////////////
   void Broadcast(int root,void* data, int bytes);
+
+  ////////////////////////////////////////////////////////////
+  // All2All down one dimension
+  ////////////////////////////////////////////////////////////
+  template<class T> void AllToAll(int dim,std::vector<T> &in, std::vector<T> &out){
+    assert(dim>=0);
+    assert(dim<_ndimension);
+    assert(in.size()==out.size());
+    int numnode = _processors[dim];
+    uint64_t bytes=sizeof(T);
+    uint64_t words=in.size()/numnode;
+    assert(numnode * words == in.size());
+    assert(words < (1ULL<<31));
+    AllToAll(dim,(void *)&in[0],(void *)&out[0],words,bytes);
+  }
+  void AllToAll(int dim  ,void *in,void *out,uint64_t words,uint64_t bytes);
+  void AllToAll(void  *in,void *out,uint64_t words         ,uint64_t bytes);
   
   template<class obj> void Broadcast(int root,obj &data)
     {

lib/communicator/Communicator_mpi.cc

@@ -1,211 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/communicator/Communicator_mpi.cc
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#include <Grid/GridCore.h>
-#include <Grid/GridQCDcore.h>
-#include <Grid/qcd/action/ActionCore.h>
-#include <mpi.h>
-
-namespace Grid {
-
-
-///////////////////////////////////////////////////////////////////////////////////////////////////
-// Info that is setup once and indept of cartesian layout
-///////////////////////////////////////////////////////////////////////////////////////////////////
-MPI_Comm CartesianCommunicator::communicator_world;
-
-// Should error check all MPI calls.
-void CartesianCommunicator::Init(int *argc, char ***argv) {
-  int flag;
-  int provided;
-  MPI_Initialized(&flag); // needed to coexist with other libs apparently
-  if ( !flag ) {
-    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
-    if ( provided != MPI_THREAD_MULTIPLE ) {
-      QCD::WilsonKernelsStatic::Comms = QCD::WilsonKernelsStatic::CommsThenCompute;
-    }
-  }
-  MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
-  ShmInitGeneric();
-}
-void CartesianCommunicator::GlobalSum(uint32_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(uint64_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalXOR(uint32_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalXOR(uint64_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(float &f){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(float *f,int N)
-{
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(double &d)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(double *d,int N)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
-{
-  int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
-  assert(ierr==0);
-}
-int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
-{
-  int rank;
-  int ierr=MPI_Cart_rank  (communicator, &coor[0], &rank);
-  assert(ierr==0);
-  return rank;
-}
-void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
-{
-  coor.resize(_ndimension);
-  int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
-  assert(ierr==0);
-}
-
-// Basic Halo comms primitive
-void CartesianCommunicator::SendToRecvFrom(void *xmit,
-					   int dest,
-					   void *recv,
-					   int from,
-					   int bytes)
-{
-  std::vector<CommsRequest_t> reqs(0);
-  SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
-  SendToRecvFromComplete(reqs);
-}
-
-void CartesianCommunicator::SendRecvPacket(void *xmit,
-					   void *recv,
-					   int sender,
-					   int receiver,
-					   int bytes)
-{
-  MPI_Status stat;
-  assert(sender != receiver);
-  int tag = sender;
-  if ( _processor == sender ) {
-    MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
-  }
-  if ( _processor == receiver ) { 
-    MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
-  }
-}
-
-// Basic Halo comms primitive
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						void *xmit,
-						int dest,
-						void *recv,
-						int from,
-						int bytes)
-{
-  int myrank = _processor;
-  int ierr;
-  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
-    MPI_Request xrq;
-    MPI_Request rrq;
-
-    ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
-    ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
-    
-    assert(ierr==0);
-    list.push_back(xrq);
-    list.push_back(rrq);
-  } else { 
-    // Give the CPU to MPI immediately; can use threads to overlap optionally
-    ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
-		      recv,bytes,MPI_CHAR,from, from,
-		      communicator,MPI_STATUS_IGNORE);
-    assert(ierr==0);
-  }
-}
-void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
-{
-  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
-    int nreq=list.size();
-    std::vector<MPI_Status> status(nreq);
-    int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-    assert(ierr==0);
-  }
-}
-
-void CartesianCommunicator::Barrier(void)
-{
-  int ierr = MPI_Barrier(communicator);
-  assert(ierr==0);
-}
-
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
-{
-  int ierr=MPI_Bcast(data,
-		     bytes,
-		     MPI_BYTE,
-		     root,
-		     communicator);
-  assert(ierr==0);
-}
-  ///////////////////////////////////////////////////////
-  // Should only be used prior to Grid Init finished.
-  // Check for this?
-  ///////////////////////////////////////////////////////
-int CartesianCommunicator::RankWorld(void){ 
-  int r; 
-  MPI_Comm_rank(communicator_world,&r);
-  return r;
-}
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
-{
-  int ierr= MPI_Bcast(data,
-		      bytes,
-		      MPI_BYTE,
-		      root,
-		      communicator_world);
-  assert(ierr==0);
-}
-
-}
-

lib/communicator/Communicator_mpi3.cc

@@ -26,89 +26,20 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
     *************************************************************************************/
     /*  END LEGAL */
 #include <Grid/GridCore.h>
-
+#include <Grid/communicator/SharedMemory.h>
-#include <mpi.h>
-
-#include <semaphore.h>
-#include <fcntl.h>
-#include <unistd.h>
-#include <limits.h>
-#include <sys/types.h>
-#include <sys/ipc.h>
-#include <sys/shm.h>
-#include <sys/mman.h>
-#include <zlib.h>
-#ifdef HAVE_NUMAIF_H
-#include <numaif.h>
-#endif
-
 
 namespace Grid {
 
-///////////////////////////////////////////////////////////////////////////////////////////////////
+Grid_MPI_Comm       CartesianCommunicator::communicator_world;
-// Info that is setup once and indept of cartesian layout
-///////////////////////////////////////////////////////////////////////////////////////////////////
-int CartesianCommunicator::ShmSetup = 0;
 
-int CartesianCommunicator::ShmRank;
+////////////////////////////////////////////
-int CartesianCommunicator::ShmSize;
+// First initialise of comms system
-int CartesianCommunicator::GroupRank;
+////////////////////////////////////////////
-int CartesianCommunicator::GroupSize;
+void CartesianCommunicator::Init(int *argc, char ***argv) 
-int CartesianCommunicator::WorldRank;
-int CartesianCommunicator::WorldSize;
-
-MPI_Comm CartesianCommunicator::communicator_world;
-MPI_Comm CartesianCommunicator::ShmComm;
-MPI_Win  CartesianCommunicator::ShmWindow;
-
-std::vector<int> CartesianCommunicator::GroupRanks;  
-std::vector<int> CartesianCommunicator::MyGroup;
-std::vector<void *> CartesianCommunicator::ShmCommBufs;
-
-int CartesianCommunicator::NodeCount(void)    { return GroupSize;};
-int CartesianCommunicator::RankCount(void)    { return WorldSize;};
-
-
-#undef FORCE_COMMS
-void *CartesianCommunicator::ShmBufferSelf(void)
 {
-  return ShmCommBufs[ShmRank];
-}
-void *CartesianCommunicator::ShmBuffer(int rank)
-{
-  int gpeer = GroupRanks[rank];
-#ifdef FORCE_COMMS
-  return NULL;
-#endif
-  if (gpeer == MPI_UNDEFINED){
-    return NULL;
-  } else { 
-    return ShmCommBufs[gpeer];
-  }
-}
-void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p)
-{
-  static int count =0;
-  int gpeer = GroupRanks[rank];
-  assert(gpeer!=ShmRank); // never send to self
-  assert(rank!=WorldRank);// never send to self
-#ifdef FORCE_COMMS
-  return NULL;
-#endif
-  if (gpeer == MPI_UNDEFINED){
-    return NULL;
-  } else { 
-    uint64_t offset = (uint64_t)local_p - (uint64_t)ShmCommBufs[ShmRank];
-    uint64_t remote = (uint64_t)ShmCommBufs[gpeer]+offset;
-    return (void *) remote;
-  }
-}
-
-void CartesianCommunicator::Init(int *argc, char ***argv) {
 
   int flag;
   int provided;
-  //  mtrace();
 
   MPI_Initialized(&flag); // needed to coexist with other libs apparently
   if ( !flag ) {
@@ -119,483 +50,213 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
   Grid_quiesce_nodes();
 
   MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
-  MPI_Comm_rank(communicator_world,&WorldRank);
-  MPI_Comm_size(communicator_world,&WorldSize);
 
-  if ( WorldRank == 0 ) {
+  GlobalSharedMemory::Init(communicator_world);
-    std::cout << GridLogMessage<< "Initialising MPI "<< WorldRank <<"/"<<WorldSize <<std::endl;
+  GlobalSharedMemory::SharedMemoryAllocate(
-  }
+		   GlobalSharedMemory::MAX_MPI_SHM_BYTES,
-
+		   GlobalSharedMemory::Hugepages);
-  /////////////////////////////////////////////////////////////////////
-  // Split into groups that can share memory
-  /////////////////////////////////////////////////////////////////////
-  MPI_Comm_split_type(communicator_world, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&ShmComm);
-  MPI_Comm_rank(ShmComm     ,&ShmRank);
-  MPI_Comm_size(ShmComm     ,&ShmSize);
-  GroupSize = WorldSize/ShmSize;
-
-  /////////////////////////////////////////////////////////////////////
-  // find world ranks in our SHM group (i.e. which ranks are on our node)
-  /////////////////////////////////////////////////////////////////////
-  MPI_Group WorldGroup, ShmGroup;
-  MPI_Comm_group (communicator_world, &WorldGroup); 
-  MPI_Comm_group (ShmComm, &ShmGroup);
-  
-  std::vector<int> world_ranks(WorldSize); 
-  GroupRanks.resize(WorldSize); 
-  for(int r=0;r<WorldSize;r++) world_ranks[r]=r;
-  
-  MPI_Group_translate_ranks (WorldGroup,WorldSize,&world_ranks[0],ShmGroup, &GroupRanks[0]); 
-
-  ///////////////////////////////////////////////////////////////////
-  // Identify who is in my group and noninate the leader
-  ///////////////////////////////////////////////////////////////////
-  int g=0;
-  MyGroup.resize(ShmSize);
-  for(int rank=0;rank<WorldSize;rank++){
-    if(GroupRanks[rank]!=MPI_UNDEFINED){
-      assert(g<ShmSize);
-      MyGroup[g++] = rank;
-    }
-  }
-  
-  std::sort(MyGroup.begin(),MyGroup.end(),std::less<int>());
-  int myleader = MyGroup[0];
-  
-  std::vector<int> leaders_1hot(WorldSize,0);
-  std::vector<int> leaders_group(GroupSize,0);
-  leaders_1hot [ myleader ] = 1;
-    
-  ///////////////////////////////////////////////////////////////////
-  // global sum leaders over comm world
-  ///////////////////////////////////////////////////////////////////
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&leaders_1hot[0],WorldSize,MPI_INT,MPI_SUM,communicator_world);
-  assert(ierr==0);
-  ///////////////////////////////////////////////////////////////////
-  // find the group leaders world rank
-  ///////////////////////////////////////////////////////////////////
-  int group=0;
-  for(int l=0;l<WorldSize;l++){
-    if(leaders_1hot[l]){
-      leaders_group[group++] = l;
-    }
-  }
-  ///////////////////////////////////////////////////////////////////
-  // Identify the rank of the group in which I (and my leader) live
-  ///////////////////////////////////////////////////////////////////
-  GroupRank=-1;
-  for(int g=0;g<GroupSize;g++){
-    if (myleader == leaders_group[g]){
-      GroupRank=g;
-    }
-  }
-  assert(GroupRank!=-1);
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // allocate the shared window for our group
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  MPI_Barrier(ShmComm);
-
-  ShmCommBuf = 0;
-  ShmCommBufs.resize(ShmSize);
-
-  ////////////////////////////////////////////////////////////////////////////////////////////
-  // Hugetlbf and others map filesystems as mappable huge pages
-  ////////////////////////////////////////////////////////////////////////////////////////////
-#ifdef GRID_MPI3_SHMMMAP
-  char shm_name [NAME_MAX];
-  for(int r=0;r<ShmSize;r++){
-    
-    size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
-    sprintf(shm_name,GRID_SHM_PATH "/Grid_mpi3_shm_%d_%d",GroupRank,r);
-    //sprintf(shm_name,"/var/lib/hugetlbfs/group/wheel/pagesize-2MB/" "Grid_mpi3_shm_%d_%d",GroupRank,r);
-    //    printf("Opening file %s \n",shm_name);
-    int fd=open(shm_name,O_RDWR|O_CREAT,0666);
-    if ( fd == -1) { 
-      printf("open %s failed\n",shm_name);
-      perror("open hugetlbfs");
-      exit(0);
-    }
-    int mmap_flag = MAP_SHARED ;
-#ifdef MAP_POPULATE    
-    mmap_flag|=MAP_POPULATE;
-#endif
-#ifdef MAP_HUGETLB
-    if ( Hugepages ) mmap_flag |= MAP_HUGETLB;
-#endif
-    void *ptr = (void *) mmap(NULL, MAX_MPI_SHM_BYTES, PROT_READ | PROT_WRITE, mmap_flag,fd, 0); 
-    if ( ptr == (void *)MAP_FAILED ) {    
-      printf("mmap %s failed\n",shm_name);
-      perror("failed mmap");      assert(0);    
-    }
-    assert(((uint64_t)ptr&0x3F)==0);
-    ShmCommBufs[r] =ptr;
-    
-  }
-#endif
-  ////////////////////////////////////////////////////////////////////////////////////////////
-  // POSIX SHMOPEN ; as far as I know Linux does not allow EXPLICIT HugePages with this case
-  // tmpfs (Larry Meadows says) does not support explicit huge page, and this is used for 
-  // the posix shm virtual file system
-  ////////////////////////////////////////////////////////////////////////////////////////////
-#ifdef GRID_MPI3_SHMOPEN
-  char shm_name [NAME_MAX];
-  if ( ShmRank == 0 ) {
-    for(int r=0;r<ShmSize;r++){
-
-      size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
-
-      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",GroupRank,r);
-
-      shm_unlink(shm_name);
-      int fd=shm_open(shm_name,O_RDWR|O_CREAT,0666);
-      if ( fd < 0 ) {	perror("failed shm_open");	assert(0);      }
-      ftruncate(fd, size);
-      
-      int mmap_flag = MAP_SHARED;
-#ifdef MAP_POPULATE 
-      mmap_flag |= MAP_POPULATE;
-#endif
-#ifdef MAP_HUGETLB
-      if (Hugepages) mmap_flag |= MAP_HUGETLB;
-#endif
-      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
-
-      if ( ptr == (void * )MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
-      assert(((uint64_t)ptr&0x3F)==0);
-
-// Experiments; Experiments; Try to force numa domain on the shm segment if we have numaif.h
-#if 0
-//#ifdef HAVE_NUMAIF_H
-	int status;
-	int flags=MPOL_MF_MOVE;
-#ifdef KNL
-	int nodes=1; // numa domain == MCDRAM
-	// Find out if in SNC2,SNC4 mode ?
-#else
-	int nodes=r; // numa domain == MPI ID
-#endif
-	unsigned long count=1;
-	for(uint64_t page=0;page<size;page+=4096){
-	  void *pages = (void *) ( page + (uint64_t)ptr );
-	  uint64_t *cow_it = (uint64_t *)pages;	*cow_it = 1;
-	  ierr= move_pages(0,count, &pages,&nodes,&status,flags);
-	  if (ierr && (page==0)) perror("numa relocate command failed");
-	}
-#endif
-	ShmCommBufs[r] =ptr;
-      
-    }
-  }
-
-  MPI_Barrier(ShmComm);
-
-  if ( ShmRank != 0 ) { 
-    for(int r=0;r<ShmSize;r++){
-      size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES ;
-    
-      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",GroupRank,r);
-
-      int fd=shm_open(shm_name,O_RDWR,0666);
-      if ( fd<0 ) {	perror("failed shm_open");	assert(0);      }
-
-      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
-      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
-      assert(((uint64_t)ptr&0x3F)==0);
-      ShmCommBufs[r] =ptr;
-    }
-  }
-#endif
-  ////////////////////////////////////////////////////////////////////////////////////////////
-  // SHMGET SHMAT and SHM_HUGETLB flag
-  ////////////////////////////////////////////////////////////////////////////////////////////
-#ifdef GRID_MPI3_SHMGET
-  std::vector<int> shmids(ShmSize);
-
-  if ( ShmRank == 0 ) {
-    for(int r=0;r<ShmSize;r++){
-      size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
-      key_t key   = IPC_PRIVATE;
-      int flags = IPC_CREAT | SHM_R | SHM_W;
-#ifdef SHM_HUGETLB
-      if (Hugepages) flags|=SHM_HUGETLB;
-#endif
-      if ((shmids[r]= shmget(key,size, flags)) ==-1) {
-	int errsv = errno;
-	printf("Errno %d\n",errsv);
-	printf("key   %d\n",key);
-	printf("size  %lld\n",size);
-	printf("flags %d\n",flags);
-	perror("shmget");
-	exit(1);
-      } else { 
-	printf("shmid: 0x%x\n", shmids[r]);
-      }
-    }
-  }
-  MPI_Barrier(ShmComm);
-  MPI_Bcast(&shmids[0],ShmSize*sizeof(int),MPI_BYTE,0,ShmComm);
-  MPI_Barrier(ShmComm);
-
-  for(int r=0;r<ShmSize;r++){
-    ShmCommBufs[r] = (uint64_t *)shmat(shmids[r], NULL,0);
-    if (ShmCommBufs[r] == (uint64_t *)-1) {
-      perror("Shared memory attach failure");
-      shmctl(shmids[r], IPC_RMID, NULL);
-      exit(2);
-    }
-    printf("shmaddr: %p\n", ShmCommBufs[r]);
-  }
-  MPI_Barrier(ShmComm);
-  // Mark for clean up
-  for(int r=0;r<ShmSize;r++){
-    shmctl(shmids[r], IPC_RMID,(struct shmid_ds *)NULL);
-  }
-  MPI_Barrier(ShmComm);
-
-#endif
-  ShmCommBuf         = ShmCommBufs[ShmRank];
-
-  MPI_Barrier(ShmComm);
-  if ( ShmRank == 0 ) {
-    for(int r=0;r<ShmSize;r++){
-      uint64_t * check = (uint64_t *) ShmCommBufs[r];
-      check[0] = GroupRank;
-      check[1] = r;
-      check[2] = 0x5A5A5A;
-    }
-  }
-
-  MPI_Barrier(ShmComm);
-  for(int r=0;r<ShmSize;r++){
-    uint64_t * check = (uint64_t *) ShmCommBufs[r];
-    
-    assert(check[0]==GroupRank);
-    assert(check[1]==r);
-    assert(check[2]==0x5A5A5A);
-
-  }
-  MPI_Barrier(ShmComm);
-
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // Verbose for now
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  if (WorldRank == 0){
-    std::cout<<GridLogMessage<< "Grid MPI-3 configuration: detected ";
-    std::cout<< WorldSize << " Ranks " ;
-    std::cout<< GroupSize << " Nodes " ;
-    std::cout<< " with "<< ShmSize  << " ranks-per-node "<<std::endl;
-    
-    std::cout<<GridLogMessage     <<"Grid MPI-3 configuration: allocated shared memory region of size ";
-    std::cout<<std::hex << MAX_MPI_SHM_BYTES <<" ShmCommBuf address = "<<ShmCommBuf << std::dec<<std::endl;
-
-    for(int g=0;g<GroupSize;g++){
-      std::cout<<GridLogMessage<<" Node "<<g<<" led by MPI rank "<<leaders_group[g]<<std::endl;
-    }
-
-    std::cout<<GridLogMessage<<" Boss Node Shm Pointers are {";
-    for(int g=0;g<ShmSize;g++){
-      std::cout<<std::hex<<ShmCommBufs[g]<<std::dec;
-      if(g!=ShmSize-1) std::cout<<",";
-      else std::cout<<"}"<<std::endl;
-    }
-  }
-  
-  for(int g=0;g<GroupSize;g++){
-    if ( (ShmRank == 0) && (GroupRank==g) )  std::cout<<GridLogMessage<<"["<<g<<"] Node Group "<<g<<" is ranks {";
-    for(int r=0;r<ShmSize;r++){
-      if ( (ShmRank == 0) && (GroupRank==g) ) {
-	std::cout<<MyGroup[r];
-	if(r<ShmSize-1) std::cout<<",";
-	else std::cout<<"}"<<std::endl<<std::flush;
-      }
-      MPI_Barrier(communicator_world);
-    }
-  }
-
-  assert(ShmSetup==0);  ShmSetup=1;
 }
 
-////////////////////////////////////////////////////////////////////////////////////////////////////////////
+///////////////////////////////////////////////////////////////////////////
-// Want to implement some magic ... Group sub-cubes into those on same node
+// Use cartesian communicators now even in MPI3
-////////////////////////////////////////////////////////////////////////////////////////////////////////////
+///////////////////////////////////////////////////////////////////////////
-void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &dest,int &source)
+void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
 {
-  std::vector<int> coor = _processor_coor; // my coord
+  int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
-  assert(std::abs(shift) <_processors[dim]);
+  assert(ierr==0);
-
+}
-  coor[dim] = (_processor_coor[dim] + shift + _processors[dim])%_processors[dim];
-  Lexicographic::IndexFromCoor(coor,source,_processors);
-  source = LexicographicToWorldRank[source];
-
-  coor[dim] = (_processor_coor[dim] - shift + _processors[dim])%_processors[dim];
-  Lexicographic::IndexFromCoor(coor,dest,_processors);
-  dest = LexicographicToWorldRank[dest];
-
-}// rank is world rank.
-
 int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
 {
   int rank;
-  Lexicographic::IndexFromCoor(coor,rank,_processors);
+  int ierr=MPI_Cart_rank  (communicator, &coor[0], &rank);
-  rank = LexicographicToWorldRank[rank];
+  assert(ierr==0);
   return rank;
-}// rank is world rank
+}
-
 void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
 {
-  int lr=-1;
+  coor.resize(_ndimension);
-  for(int r=0;r<WorldSize;r++){// map world Rank to lexico and then to coor
+  int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
-    if( LexicographicToWorldRank[r]==rank) lr = r;
+  assert(ierr==0);
-  }
+}
-  assert(lr!=-1);
+
-  Lexicographic::CoorFromIndex(coor,lr,_processors);
+////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Initialises from communicator_world
+////////////////////////////////////////////////////////////////////////////////////////////////////////
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors) 
+{
+  MPI_Comm optimal_comm;
+  GlobalSharedMemory::OptimalCommunicator    (processors,optimal_comm); // Remap using the shared memory optimising routine
+  InitFromMPICommunicator(processors,optimal_comm);
+  SetCommunicator(optimal_comm);
 }
 
 //////////////////////////////////
 // Try to subdivide communicator
 //////////////////////////////////
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent) 
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank)    
-  : CartesianCommunicator(processors) 
 {
-  std::cout << "Attempts to split MPI3 communicators will fail until implemented" <<std::endl;
-}
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
-{ 
-  int ierr;
-  communicator=communicator_world;
-
   _ndimension = processors.size();
 
+  int parent_ndimension = parent._ndimension; assert(_ndimension >= parent._ndimension);
+  std::vector<int> parent_processor_coor(_ndimension,0);
+  std::vector<int> parent_processors    (_ndimension,1);
+
+  // Can make 5d grid from 4d etc...
+  int pad = _ndimension-parent_ndimension;
+  for(int d=0;d<parent_ndimension;d++){
+    parent_processor_coor[pad+d]=parent._processor_coor[d];
+    parent_processors    [pad+d]=parent._processors[d];
+  }
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // split the communicator
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  //  int Nparent = parent._processors ; 
+  //  std::cout << " splitting from communicator "<<parent.communicator <<std::endl;
+  int Nparent;
+  MPI_Comm_size(parent.communicator,&Nparent);
+  //  std::cout << " Parent size  "<<Nparent <<std::endl;
+
+  int childsize=1;
+  for(int d=0;d<processors.size();d++) {
+    childsize *= processors[d];
+  }
+  int Nchild = Nparent/childsize;
+  assert (childsize * Nchild == Nparent);
+
+  //  std::cout << " child size  "<<childsize <<std::endl;
+
+  std::vector<int> ccoor(_ndimension); // coor within subcommunicator
+  std::vector<int> scoor(_ndimension); // coor of split within parent
+  std::vector<int> ssize(_ndimension); // coor of split within parent
+
+  for(int d=0;d<_ndimension;d++){
+    ccoor[d] = parent_processor_coor[d] % processors[d];
+    scoor[d] = parent_processor_coor[d] / processors[d];
+    ssize[d] = parent_processors[d]     / processors[d];
+  }
+
+  // rank within subcomm ; srank is rank of subcomm within blocks of subcomms
+  int crank;  
+  // Mpi uses the reverse Lexico convention to us; so reversed routines called
+  Lexicographic::IndexFromCoorReversed(ccoor,crank,processors); // processors is the split grid dimensions
+  Lexicographic::IndexFromCoorReversed(scoor,srank,ssize);      // ssize is the number of split grids
+
+  MPI_Comm comm_split;
+  if ( Nchild > 1 ) { 
+
+    if(0){
+      std::cout << GridLogMessage<<"Child communicator of "<< std::hex << parent.communicator << std::dec<<std::endl;
+      std::cout << GridLogMessage<<" parent grid["<< parent._ndimension<<"]    ";
+      for(int d=0;d<parent._ndimension;d++)  std::cout << parent._processors[d] << " ";
+      std::cout<<std::endl;
+      
+      std::cout << GridLogMessage<<" child grid["<< _ndimension <<"]    ";
+      for(int d=0;d<processors.size();d++)  std::cout << processors[d] << " ";
+      std::cout<<std::endl;
+      
+      std::cout << GridLogMessage<<" old rank "<< parent._processor<<" coor ["<< parent._ndimension <<"]    ";
+      for(int d=0;d<parent._ndimension;d++)  std::cout << parent._processor_coor[d] << " ";
+      std::cout<<std::endl;
+      
+      std::cout << GridLogMessage<<" new split "<< srank<<" scoor ["<< _ndimension <<"]    ";
+      for(int d=0;d<processors.size();d++)  std::cout << scoor[d] << " ";
+      std::cout<<std::endl;
+      
+      std::cout << GridLogMessage<<" new rank "<< crank<<" coor ["<< _ndimension <<"]    ";
+      for(int d=0;d<processors.size();d++)  std::cout << ccoor[d] << " ";
+      std::cout<<std::endl;
+
+      //////////////////////////////////////////////////////////////////////////////////////////////////////
+      // Declare victory
+      //////////////////////////////////////////////////////////////////////////////////////////////////////
+      std::cout << GridLogMessage<<"Divided communicator "<< parent._Nprocessors<<" into "
+		<< Nchild <<" communicators with " << childsize << " ranks"<<std::endl;
+      std::cout << " Split communicator " <<comm_split <<std::endl;
+    }
+
+    ////////////////////////////////////////////////////////////////
+    // Split the communicator
+    ////////////////////////////////////////////////////////////////
+    int ierr= MPI_Comm_split(parent.communicator,srank,crank,&comm_split);
+    assert(ierr==0);
+
+  } else {
+    srank = 0;
+    comm_split    = parent.communicator;
+    //    std::cout << " Inherited communicator " <<comm_split <<std::endl;
+  }
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Set up from the new split communicator
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  InitFromMPICommunicator(processors,comm_split);
+
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  // Take the right SHM buffers
+  //////////////////////////////////////////////////////////////////////////////////////////////////////
+  SetCommunicator(comm_split);
+
+  if(0){ 
+    std::cout << " ndim " <<_ndimension<<" " << parent._ndimension << std::endl;
+    for(int d=0;d<processors.size();d++){
+      std::cout << d<< " " << _processor_coor[d] <<" " <<  ccoor[d]<<std::endl;
+    }
+  }
+  for(int d=0;d<processors.size();d++){
+    assert(_processor_coor[d] == ccoor[d] );
+  }
+}
+
+void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base)
+{
+  _ndimension = processors.size();
+  _processor_coor.resize(_ndimension);
+
+  /////////////////////////////////
+  // Count the requested nodes
+  /////////////////////////////////
+  _Nprocessors=1;
+  _processors = processors;
+  for(int i=0;i<_ndimension;i++){
+    _Nprocessors*=_processors[i];
+  }
+
+  std::vector<int> periodic(_ndimension,1);
+  MPI_Cart_create(communicator_base, _ndimension,&_processors[0],&periodic[0],0,&communicator);
+  MPI_Comm_rank(communicator,&_processor);
+  MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
+
+  if ( 0 && (communicator_base != communicator_world) ) {
+    std::cout << "InitFromMPICommunicator Cartesian communicator created with a non-world communicator"<<std::endl;
+    std::cout << " new communicator rank "<<_processor<< " coor ["<<_ndimension<<"] ";
+    for(int d=0;d<_processors.size();d++){
+      std::cout << _processor_coor[d]<<" ";
+    }
+    std::cout << std::endl;
+  }
+
+  int Size;
+  MPI_Comm_size(communicator,&Size);
+
   communicator_halo.resize (2*_ndimension);
   for(int i=0;i<_ndimension*2;i++){
     MPI_Comm_dup(communicator,&communicator_halo[i]);
   }
+  assert(Size==_Nprocessors);
+}
 
-  ////////////////////////////////////////////////////////////////
+CartesianCommunicator::~CartesianCommunicator()
-  // Assert power of two shm_size.
+{
-  ////////////////////////////////////////////////////////////////
+  int MPI_is_finalised;
-  int log2size = -1;
+  MPI_Finalized(&MPI_is_finalised);
-  for(int i=0;i<=MAXLOG2RANKSPERNODE;i++){  
+  if (communicator && !MPI_is_finalised) {
-    if ( (0x1<<i) == ShmSize ) {
+    MPI_Comm_free(&communicator);
-      log2size = i;
+    for(int i=0;i<communicator_halo.size();i++){
-      break;
+      MPI_Comm_free(&communicator_halo[i]);
     }
-  }
+  }  
-  assert(log2size != -1);
+}
-
-  ////////////////////////////////////////////////////////////////
-  // Identify subblock of ranks on node spreading across dims
-  // in a maximally symmetrical way
-  ////////////////////////////////////////////////////////////////
-  std::vector<int> WorldDims = processors;
-
-  ShmDims.resize  (_ndimension,1);
-  GroupDims.resize(_ndimension);
-  ShmCoor.resize  (_ndimension);
-  GroupCoor.resize(_ndimension);
-  WorldCoor.resize(_ndimension);
-
-  int dim = 0;
-  for(int l2=0;l2<log2size;l2++){
-    while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%_ndimension;
-    ShmDims[dim]*=2;
-    dim=(dim+1)%_ndimension;
-  }
-
-  ////////////////////////////////////////////////////////////////
-  // Establish torus of processes and nodes with sub-blockings
-  ////////////////////////////////////////////////////////////////
-  for(int d=0;d<_ndimension;d++){
-    GroupDims[d] = WorldDims[d]/ShmDims[d];
-  }
-
-  ////////////////////////////////////////////////////////////////
-  // Verbose
-  ////////////////////////////////////////////////////////////////
-#if 0
-  std::cout<< GridLogMessage << "MPI-3 usage "<<std::endl;
-  std::cout<< GridLogMessage << "SHM   ";
-  for(int d=0;d<_ndimension;d++){
-    std::cout<< ShmDims[d] <<" ";
-  }
-  std::cout<< std::endl;
-
-  std::cout<< GridLogMessage << "Group ";
-  for(int d=0;d<_ndimension;d++){
-    std::cout<< GroupDims[d] <<" ";
-  }
-  std::cout<< std::endl;
-
-  std::cout<< GridLogMessage<<"World ";
-  for(int d=0;d<_ndimension;d++){
-    std::cout<< WorldDims[d] <<" ";
-  }
-  std::cout<< std::endl;
-#endif
-  ////////////////////////////////////////////////////////////////
-  // Check processor counts match
-  ////////////////////////////////////////////////////////////////
-  _Nprocessors=1;
-  _processors = processors;
-  _processor_coor.resize(_ndimension);
-  for(int i=0;i<_ndimension;i++){
-    _Nprocessors*=_processors[i];
-  }
-  assert(WorldSize==_Nprocessors);
-      
-  ////////////////////////////////////////////////////////////////
-  // Establish mapping between lexico physics coord and WorldRank
-  ////////////////////////////////////////////////////////////////
-  Lexicographic::CoorFromIndex(GroupCoor,GroupRank,GroupDims);
-  Lexicographic::CoorFromIndex(ShmCoor,ShmRank,ShmDims);
-  for(int d=0;d<_ndimension;d++){
-    WorldCoor[d] = GroupCoor[d]*ShmDims[d]+ShmCoor[d];
-  }
-  _processor_coor = WorldCoor;
-  _processor      = WorldRank;
-
-  ///////////////////////////////////////////////////////////////////
-  // global sum Lexico to World mapping
-  ///////////////////////////////////////////////////////////////////
-  int lexico;
-  LexicographicToWorldRank.resize(WorldSize,0);
-  Lexicographic::IndexFromCoor(WorldCoor,lexico,WorldDims);
-  LexicographicToWorldRank[lexico] = WorldRank;
-  ierr=MPI_Allreduce(MPI_IN_PLACE,&LexicographicToWorldRank[0],WorldSize,MPI_INT,MPI_SUM,communicator);
-  assert(ierr==0);
-
-  for(int i=0;i<WorldSize;i++){
-
-    int wr = LexicographicToWorldRank[i];
-    //    int wr = i;
-
-    std::vector<int> coor(_ndimension);
-    ProcessorCoorFromRank(wr,coor); // from world rank
-    int ck = RankFromProcessorCoor(coor);
-    assert(ck==wr);
-
-    if ( wr == WorldRank ) { 
-      for(int j=0;j<coor.size();j++) {
-	assert(coor[j] == _processor_coor[j]);
-      }
-    }
-    /*
-    std::cout << GridLogMessage<< " Lexicographic "<<i;
-    std::cout << " MPI rank      "<<wr;
-    std::cout << " Coor          ";
-    for(int j=0;j<coor.size();j++) std::cout << coor[j];
-    std::cout<< std::endl;
-    */
-    /////////////////////////////////////////////////////
-    // Check everyone agrees on everyone elses coords
-    /////////////////////////////////////////////////////
-    std::vector<int> mcoor = coor;
-    this->Broadcast(0,(void *)&mcoor[0],mcoor.size()*sizeof(int));
-    for(int d = 0 ; d< _ndimension; d++) {
-      assert(coor[d] == mcoor[d]);
-    }
-  }
-};
 void CartesianCommunicator::GlobalSum(uint32_t &u){
   int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
   assert(ierr==0);
@@ -712,34 +373,31 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
 							 int from,
 							 int bytes,int dir)
 {
-  assert(dir < communicator_halo.size());
+  int ncomm  =communicator_halo.size(); 
+  int commdir=dir%ncomm;
 
   MPI_Request xrq;
   MPI_Request rrq;
 
   int ierr;
-  int gdest = GroupRanks[dest];
+  int gdest = ShmRanks[dest];
-  int gfrom = GroupRanks[from];
+  int gfrom = ShmRanks[from];
-  int gme   = GroupRanks[_processor];
+  int gme   = ShmRanks[_processor];
 
   assert(dest != _processor);
   assert(from != _processor);
   assert(gme  == ShmRank);
   double off_node_bytes=0.0;
 
-#ifdef FORCE_COMMS
-  gdest = MPI_UNDEFINED;
-  gfrom = MPI_UNDEFINED;
-#endif
   if ( gfrom ==MPI_UNDEFINED) {
-    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[dir],&rrq);
+    ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[commdir],&rrq);
     assert(ierr==0);
     list.push_back(rrq);
     off_node_bytes+=bytes;
   }
 
   if ( gdest == MPI_UNDEFINED ) {
-    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[dir],&xrq);
+    ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[commdir],&xrq);
     assert(ierr==0);
     list.push_back(xrq);
     off_node_bytes+=bytes;
@@ -799,5 +457,38 @@ void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
   assert(ierr==0);
 }
 
+void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  std::vector<int> row(_ndimension,1);
+  assert(dim>=0 && dim<_ndimension);
+
+  //  Split the communicator
+  row[dim] = _processors[dim];
+
+  int me;
+  CartesianCommunicator Comm(row,*this,me);
+  Comm.AllToAll(in,out,words,bytes);
+}
+void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  // MPI is a pain and uses "int" arguments
+  // 64*64*64*128*16 == 500Million elements of data.
+  // When 24*4 bytes multiples get 50x 10^9 >>> 2x10^9 Y2K bug.
+  // (Turns up on 32^3 x 64 Gparity too)
+  MPI_Datatype object;
+  int iwords; 
+  int ibytes;
+  iwords = words;
+  ibytes = bytes;
+  assert(words == iwords); // safe to cast to int ?
+  assert(bytes == ibytes); // safe to cast to int ?
+  MPI_Type_contiguous(ibytes,MPI_BYTE,&object);
+  MPI_Type_commit(&object);
+  MPI_Alltoall(in,iwords,object,out,iwords,object,communicator);
+  MPI_Type_free(&object);
+}
+
+
+
 }
 

lib/communicator/Communicator_mpi3_leader.cc

@@ -1,988 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/communicator/Communicator_mpi.cc
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#include "Grid.h"
-#include <mpi.h>
-//#include <numaif.h>
-
-////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-/// Workarounds:
-/// i) bloody mac os doesn't implement unnamed semaphores since it is "optional" posix.
-///    darwin dispatch semaphores don't seem to be multiprocess.
-///
-/// ii) openmpi under --mca shmem posix works with two squadrons per node; 
-///     openmpi under default mca settings (I think --mca shmem mmap) on MacOS makes two squadrons map the SAME
-///     memory as each other, despite their living on different communicators. This appears to be a bug in OpenMPI.
-///
-////////////////////////////////////////////////////////////////////////////////////////////////////////////////
-#include <semaphore.h>
-#include <fcntl.h>
-#include <unistd.h>
-#include <limits.h>
-typedef sem_t *Grid_semaphore;
-
-
-#error  /*THis is deprecated*/
-
-#if 0 
-#define SEM_INIT(S)      S = sem_open(sem_name,0,0600,0); assert ( S != SEM_FAILED );
-#define SEM_INIT_EXCL(S) sem_unlink(sem_name); S = sem_open(sem_name,O_CREAT|O_EXCL,0600,0); assert ( S != SEM_FAILED );
-#define SEM_POST(S) assert ( sem_post(S) == 0 ); 
-#define SEM_WAIT(S) assert ( sem_wait(S) == 0 );
-#else
-#define SEM_INIT(S)      ;
-#define SEM_INIT_EXCL(S) ;
-#define SEM_POST(S) ;
-#define SEM_WAIT(S) ;
-#endif
-#include <sys/mman.h>
-
-namespace Grid {
-
-enum { COMMAND_ISEND, COMMAND_IRECV, COMMAND_WAITALL, COMMAND_SENDRECV };
-
-struct Descriptor {
-  uint64_t buf;
-  size_t bytes;
-  int rank;
-  int tag;
-  int command;
-  uint64_t xbuf;
-  uint64_t rbuf;
-  int xtag;
-  int rtag;
-  int src;
-  int dest;
-  MPI_Request request;
-};
-
-const int pool = 48;
-
-class SlaveState {
-public:
-  volatile int head;
-  volatile int start;
-  volatile int tail;
-  volatile Descriptor Descrs[pool];
-};
-
-class Slave {
-public:
-  Grid_semaphore  sem_head;
-  Grid_semaphore  sem_tail;
-  SlaveState *state;
-  MPI_Comm squadron;
-  uint64_t     base;
-  int universe_rank;
-  int vertical_rank;
-  char sem_name [NAME_MAX];
-  ////////////////////////////////////////////////////////////
-  // Descriptor circular pointers
-  ////////////////////////////////////////////////////////////
-  Slave() {};
-
-  void Init(SlaveState * _state,MPI_Comm _squadron,int _universe_rank,int _vertical_rank);
-
-  void SemInit(void) {
-    sprintf(sem_name,"/Grid_mpi3_sem_head_%d",universe_rank);
-    SEM_INIT(sem_head);
-    sprintf(sem_name,"/Grid_mpi3_sem_tail_%d",universe_rank);
-    SEM_INIT(sem_tail);
-  }  
-  void SemInitExcl(void) {
-    sprintf(sem_name,"/Grid_mpi3_sem_head_%d",universe_rank);
-    SEM_INIT_EXCL(sem_head);
-    sprintf(sem_name,"/Grid_mpi3_sem_tail_%d",universe_rank);
-    SEM_INIT_EXCL(sem_tail);
-  }  
-  void WakeUpDMA(void) { 
-    SEM_POST(sem_head);
-  };
-  void WakeUpCompute(void) { 
-    SEM_POST(sem_tail);
-  };
-  void WaitForCommand(void) { 
-    SEM_WAIT(sem_head);
-  };
-  void WaitForComplete(void) { 
-    SEM_WAIT(sem_tail);
-  };
-  void EventLoop (void) {
-    //    std::cout<< " Entering event loop "<<std::endl;
-    while(1){
-      WaitForCommand();
-      //      std::cout << "Getting command "<<std::endl;
-#if 0
-      _mm_monitor((void *)&state->head,0,0);
-      int s=state->start;
-      if ( s != state->head ) {
-	_mm_mwait(0,0);
-      }
-#endif
-      Event();
-    }
-  }
-
-  int Event (void) ;
-
-  uint64_t QueueCommand(int command,void *buf, int bytes, int hashtag, MPI_Comm comm,int u_rank) ;
-  void QueueSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) ;
-
-  void WaitAll() {
-    //    std::cout << "Queueing WAIT command  "<<std::endl;
-    QueueCommand(COMMAND_WAITALL,0,0,0,squadron,0);
-    //    std::cout << "Waking up DMA "<<std::endl;
-    WakeUpDMA();
-    //    std::cout << "Waiting from semaphore "<<std::endl;
-    WaitForComplete();
-    //    std::cout << "Checking FIFO is empty "<<std::endl;
-    while ( state->tail != state->head );
-  }
-};
-
-////////////////////////////////////////////////////////////////////////
-// One instance of a data mover.
-// Master and Slave must agree on location in shared memory
-////////////////////////////////////////////////////////////////////////
-
-class MPIoffloadEngine { 
-public:
-
-  static std::vector<Slave> Slaves;
-
-  static int ShmSetup;
-  
-  static int UniverseRank;
-  static int UniverseSize;
-  
-  static MPI_Comm communicator_universe;
-  static MPI_Comm communicator_cached;
-
-  static MPI_Comm HorizontalComm;
-  static int HorizontalRank;
-  static int HorizontalSize;
-  
-  static MPI_Comm VerticalComm;
-  static MPI_Win  VerticalWindow; 
-  static int VerticalSize;
-  static int VerticalRank;
-  
-  static std::vector<void *> VerticalShmBufs;
-  static std::vector<std::vector<int> > UniverseRanks;
-  static std::vector<int> UserCommunicatorToWorldRanks; 
-  
-  static MPI_Group WorldGroup, CachedGroup;
-  
-  static void CommunicatorInit (MPI_Comm &communicator_world,
-				MPI_Comm &ShmComm,
-				void * &ShmCommBuf);
-
-  static void MapCommRankToWorldRank(int &hashtag, int & comm_world_peer,int tag, MPI_Comm comm,int commrank);
-
-  /////////////////////////////////////////////////////////
-  // routines for master proc must handle any communicator
-  /////////////////////////////////////////////////////////
-
-  static void QueueSend(int slave,void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
-     //    std::cout<< " Queueing send  "<< bytes<< " slave "<< slave << " to comm "<<rank  <<std::endl;
-    Slaves[slave].QueueCommand(COMMAND_ISEND,buf,bytes,tag,comm,rank);
-    //    std::cout << "Queued send command to rank "<< rank<< " via "<<slave <<std::endl;
-    Slaves[slave].WakeUpDMA();
-    //    std::cout << "Waking up DMA "<< slave<<std::endl;
-  };
-
-  static void QueueSendRecv(int slave,void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) 
-  {
-    Slaves[slave].QueueSendRecv(xbuf,rbuf,bytes,xtag,rtag,comm,dest,src);
-    Slaves[slave].WakeUpDMA();
-  }
-
-  static void QueueRecv(int slave, void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
-    //    std::cout<< " Queueing recv "<< bytes<< " slave "<< slave << " from comm "<<rank  <<std::endl;
-    Slaves[slave].QueueCommand(COMMAND_IRECV,buf,bytes,tag,comm,rank);
-    //    std::cout << "Queued recv command from rank "<< rank<< " via "<<slave <<std::endl;
-    Slaves[slave].WakeUpDMA();
-    //    std::cout << "Waking up DMA "<< slave<<std::endl;
-  };
-
-  static void WaitAll() {
-    for(int s=1;s<VerticalSize;s++) {
-      //      std::cout << "Waiting for slave "<< s<<std::endl;
-      Slaves[s].WaitAll();
-    }
-    //    std::cout << " Wait all Complete "<<std::endl;
-  };
-
-  static void GetWork(int nwork, int me, int & mywork, int & myoff,int units){
-    int basework = nwork/units;
-    int backfill = units-(nwork%units);
-    if ( me >= units ) { 
-      mywork = myoff = 0;
-    } else { 
-      mywork = (nwork+me)/units;
-      myoff  = basework * me;
-      if ( me > backfill ) 
-	myoff+= (me-backfill);
-    }
-    return;
-  };
-
-  static void QueueRoundRobinSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) {
-    uint8_t * cxbuf = (uint8_t *) xbuf;
-    uint8_t * crbuf = (uint8_t *) rbuf;
-    static int rrp=0;
-    int procs = VerticalSize-1;
-    int myoff=0;
-    int mywork=bytes;
-    QueueSendRecv(rrp+1,&cxbuf[myoff],&crbuf[myoff],mywork,xtag,rtag,comm,dest,src);
-    rrp = rrp+1;
-    if ( rrp == (VerticalSize-1) ) rrp = 0;
-  }
-
-  static void QueueMultiplexedSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) {
-    uint8_t * cxbuf = (uint8_t *) xbuf;
-    uint8_t * crbuf = (uint8_t *) rbuf;
-    int mywork, myoff, procs;
-    procs = VerticalSize-1;
-    for(int s=0;s<procs;s++) {
-      GetWork(bytes,s,mywork,myoff,procs);
-      QueueSendRecv(s+1,&cxbuf[myoff],&crbuf[myoff],mywork,xtag,rtag,comm,dest,src);
-    }
-  };
-  static void QueueMultiplexedSend(void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
-    uint8_t * cbuf = (uint8_t *) buf;
-    int mywork, myoff, procs;
-    procs = VerticalSize-1;
-    for(int s=0;s<procs;s++) {
-      GetWork(bytes,s,mywork,myoff,procs);
-      QueueSend(s+1,&cbuf[myoff],mywork,tag,comm,rank);
-    }
-  };
-
-  static void QueueMultiplexedRecv(void *buf, int bytes, int tag, MPI_Comm comm,int rank) {
-    uint8_t * cbuf = (uint8_t *) buf;
-    int mywork, myoff, procs;
-    procs = VerticalSize-1;
-    for(int s=0;s<procs;s++) {
-      GetWork(bytes,s,mywork,myoff,procs);
-      QueueRecv(s+1,&cbuf[myoff],mywork,tag,comm,rank);
-    }
-  };
-
-};
-
-
-///////////////////////////////////////////////////////////////////////////////////////////////////
-// Info that is setup once and indept of cartesian layout
-///////////////////////////////////////////////////////////////////////////////////////////////////
-
-std::vector<Slave> MPIoffloadEngine::Slaves;
-    
-int MPIoffloadEngine::UniverseRank;
-int MPIoffloadEngine::UniverseSize;
-
-MPI_Comm  MPIoffloadEngine::communicator_universe;
-MPI_Comm  MPIoffloadEngine::communicator_cached;
-MPI_Group MPIoffloadEngine::WorldGroup;
-MPI_Group MPIoffloadEngine::CachedGroup;
-
-MPI_Comm MPIoffloadEngine::HorizontalComm;
-int      MPIoffloadEngine::HorizontalRank;
-int      MPIoffloadEngine::HorizontalSize;
-
-MPI_Comm MPIoffloadEngine::VerticalComm;
-int      MPIoffloadEngine::VerticalSize;
-int      MPIoffloadEngine::VerticalRank;
-MPI_Win  MPIoffloadEngine::VerticalWindow; 
-std::vector<void *>            MPIoffloadEngine::VerticalShmBufs;
-std::vector<std::vector<int> > MPIoffloadEngine::UniverseRanks;
-std::vector<int>               MPIoffloadEngine::UserCommunicatorToWorldRanks; 
-
-int CartesianCommunicator::NodeCount(void)    { return HorizontalSize;};
-int MPIoffloadEngine::ShmSetup = 0;
-
-void MPIoffloadEngine::CommunicatorInit (MPI_Comm &communicator_world,
-					 MPI_Comm &ShmComm,
-					 void * &ShmCommBuf)
-{      
-  int flag;
-  assert(ShmSetup==0);  
-  
-  //////////////////////////////////////////////////////////////////////
-  // Universe is all nodes prior to squadron grouping
-  //////////////////////////////////////////////////////////////////////
-  MPI_Comm_dup (MPI_COMM_WORLD,&communicator_universe);
-  MPI_Comm_rank(communicator_universe,&UniverseRank);
-  MPI_Comm_size(communicator_universe,&UniverseSize);
-  
-  /////////////////////////////////////////////////////////////////////
-  // Split into groups that can share memory (Verticals)
-  /////////////////////////////////////////////////////////////////////
-#undef MPI_SHARED_MEM_DEBUG
-#ifdef  MPI_SHARED_MEM_DEBUG
-  MPI_Comm_split(communicator_universe,(UniverseRank/4),UniverseRank,&VerticalComm);
-#else 
-  MPI_Comm_split_type(communicator_universe, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&VerticalComm);
-#endif
-  MPI_Comm_rank(VerticalComm     ,&VerticalRank);
-  MPI_Comm_size(VerticalComm     ,&VerticalSize);
-  
-  //////////////////////////////////////////////////////////////////////
-  // Split into horizontal groups by rank in squadron
-  //////////////////////////////////////////////////////////////////////
-  MPI_Comm_split(communicator_universe,VerticalRank,UniverseRank,&HorizontalComm);
-  MPI_Comm_rank(HorizontalComm,&HorizontalRank);
-  MPI_Comm_size(HorizontalComm,&HorizontalSize);
-  assert(HorizontalSize*VerticalSize==UniverseSize);
-  
-  ////////////////////////////////////////////////////////////////////////////////
-  // What is my place in the world
-  ////////////////////////////////////////////////////////////////////////////////
-  int WorldRank=0;
-  if(VerticalRank==0) WorldRank = HorizontalRank;
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&WorldRank,1,MPI_INT,MPI_SUM,VerticalComm);
-  assert(ierr==0);
-  
-  ////////////////////////////////////////////////////////////////////////////////
-  // Where is the world in the universe?
-  ////////////////////////////////////////////////////////////////////////////////
-  UniverseRanks = std::vector<std::vector<int> >(HorizontalSize,std::vector<int>(VerticalSize,0));
-  UniverseRanks[WorldRank][VerticalRank] = UniverseRank;
-  for(int w=0;w<HorizontalSize;w++){
-    ierr=MPI_Allreduce(MPI_IN_PLACE,&UniverseRanks[w][0],VerticalSize,MPI_INT,MPI_SUM,communicator_universe);
-    assert(ierr==0);
-  }
-  
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  // allocate the shared window for our group, pass back Shm info to CartesianCommunicator
-  //////////////////////////////////////////////////////////////////////////////////////////////////////////
-  VerticalShmBufs.resize(VerticalSize);
-
-#undef MPI_SHARED_MEM
-#ifdef MPI_SHARED_MEM
-  ierr = MPI_Win_allocate_shared(CartesianCommunicator::MAX_MPI_SHM_BYTES,1,MPI_INFO_NULL,VerticalComm,&ShmCommBuf,&VerticalWindow);
-  ierr|= MPI_Win_lock_all (MPI_MODE_NOCHECK, VerticalWindow);
-  assert(ierr==0);
-  //  std::cout<<"SHM "<<ShmCommBuf<<std::endl;
-
-  for(int r=0;r<VerticalSize;r++){
-    MPI_Aint sz;
-    int dsp_unit;
-    MPI_Win_shared_query (VerticalWindow, r, &sz, &dsp_unit, &VerticalShmBufs[r]);
-    //    std::cout<<"SHM "<<r<<" " <<VerticalShmBufs[r]<<std::endl;
-  }
-#else 
-  char shm_name [NAME_MAX];
-  MPI_Barrier(VerticalComm);
-
-  if ( VerticalRank == 0 ) {
-    for(int r=0;r<VerticalSize;r++){
-
-      size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
-      if ( r>0 ) size = sizeof(SlaveState);
-
-      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldRank,r);
-      
-      shm_unlink(shm_name);
-
-      int fd=shm_open(shm_name,O_RDWR|O_CREAT,0600);
-      if ( fd < 0 ) {
-	perror("failed shm_open");
-	assert(0);
-      }
-
-      ftruncate(fd, size);
-
-      VerticalShmBufs[r] = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
-      if ( VerticalShmBufs[r] == MAP_FAILED ) { 
-	perror("failed mmap");
-	assert(0);
-      }
-
-      /*
-      for(uint64_t page=0;page<size;page+=4096){
-	void *pages = (void *) ( page + (uint64_t)VerticalShmBufs[r] );
-	int status;
-	int flags=MPOL_MF_MOVE_ALL;
-	int nodes=1; // numa domain == MCDRAM
-	unsigned long count=1;
-	ierr= move_pages(0,count, &pages,&nodes,&status,flags);
-	if (ierr && (page==0)) perror("numa relocate command failed");
-      }
-      */
-      uint64_t * check = (uint64_t *) VerticalShmBufs[r];
-      check[0] = WorldRank;
-      check[1] = r;
-
-      //      std::cout<<"SHM "<<r<<" " <<VerticalShmBufs[r]<<std::endl;
-    }
-  }
-
-  MPI_Barrier(VerticalComm);
-
-  if ( VerticalRank != 0 ) { 
-  for(int r=0;r<VerticalSize;r++){
-
-    size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES ;
-    if ( r>0 ) size = sizeof(SlaveState);
-    
-    sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldRank,r);
-    
-    int fd=shm_open(shm_name,O_RDWR|O_CREAT,0600);
-    if ( fd<0 ) {
-      perror("failed shm_open");
-      assert(0);
-    }
-    VerticalShmBufs[r] = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
-
-    uint64_t * check = (uint64_t *) VerticalShmBufs[r];
-    assert(check[0]== WorldRank);
-    assert(check[1]== r);
-    //    std::cerr<<"SHM "<<r<<" " <<VerticalShmBufs[r]<<std::endl;
-  }
-  }
-#endif
-  MPI_Barrier(VerticalComm);
-
-  //////////////////////////////////////////////////////////////////////
-  // Map rank of leader on node in their in new world, to the
-  // rank in this vertical plane's horizontal communicator
-  //////////////////////////////////////////////////////////////////////
-  communicator_world = HorizontalComm;
-  ShmComm            = VerticalComm;
-  ShmCommBuf         = VerticalShmBufs[0];
-  MPI_Comm_group (communicator_world, &WorldGroup); 
-  
-  ///////////////////////////////////////////////////////////
-  // Start the slave data movers
-  ///////////////////////////////////////////////////////////
-  if ( VerticalRank != 0 ) {
-    Slave indentured;
-    indentured.Init( (SlaveState *) VerticalShmBufs[VerticalRank], VerticalComm, UniverseRank,VerticalRank);
-    indentured.SemInitExcl();// init semaphore in shared memory
-    MPI_Barrier(VerticalComm);
-    MPI_Barrier(VerticalComm);
-    indentured.EventLoop();
-    assert(0);
-  } else {
-    Slaves.resize(VerticalSize);
-    for(int i=1;i<VerticalSize;i++){
-      Slaves[i].Init((SlaveState *)VerticalShmBufs[i],VerticalComm, UniverseRanks[HorizontalRank][i],i);
-    }
-    MPI_Barrier(VerticalComm);
-    for(int i=1;i<VerticalSize;i++){
-      Slaves[i].SemInit();// init semaphore in shared memory
-    }
-    MPI_Barrier(VerticalComm);
-  }
-  
-  ///////////////////////////////////////////////////////////
-  // Verbose for now
-  ///////////////////////////////////////////////////////////
-  
-  ShmSetup=1;
-  
-  if (UniverseRank == 0){
-      
-    std::cout<<GridLogMessage << "Grid MPI-3 configuration: detected ";
-    std::cout<<UniverseSize   << " Ranks " ;
-    std::cout<<HorizontalSize << " Nodes " ;
-    std::cout<<VerticalSize   << " with ranks-per-node "<<std::endl;
-    
-    std::cout<<GridLogMessage << "Grid MPI-3 configuration: using one lead process per node " << std::endl;
-    std::cout<<GridLogMessage << "Grid MPI-3 configuration: reduced communicator has size " << HorizontalSize << std::endl;
-    
-    for(int g=0;g<HorizontalSize;g++){
-      std::cout<<GridLogMessage<<" Node "<<g<<" led by MPI rank "<< UniverseRanks[g][0]<<std::endl;
-    }
-    
-    for(int g=0;g<HorizontalSize;g++){
-      std::cout<<GridLogMessage<<" { ";
-      for(int s=0;s<VerticalSize;s++){
-	std::cout<< UniverseRanks[g][s];
-	if ( s<VerticalSize-1 ) {
-	  std::cout<<",";
-	}
-      }
-      std::cout<<" } "<<std::endl;
-    }
-  }
-};
-
-  ///////////////////////////////////////////////////////////////////////////////////////////////
-  // Map the communicator into communicator_world, and find the neighbour.
-  // Cache the mappings; cache size is 1.
-  ///////////////////////////////////////////////////////////////////////////////////////////////
-void MPIoffloadEngine::MapCommRankToWorldRank(int &hashtag, int & comm_world_peer,int tag, MPI_Comm comm,int rank) {
-
-  if ( comm == HorizontalComm ) {
-    comm_world_peer = rank;
-    //    std::cout << " MapCommRankToWorldRank  horiz " <<rank<<"->"<<comm_world_peer<<std::endl;
-  } else if ( comm == communicator_cached ) {
-    comm_world_peer = UserCommunicatorToWorldRanks[rank];
-    //    std::cout << " MapCommRankToWorldRank  cached " <<rank<<"->"<<comm_world_peer<<std::endl;
-  } else { 
-    
-    int size;
-
-    MPI_Comm_size(comm,&size);
-
-    UserCommunicatorToWorldRanks.resize(size);
-
-    std::vector<int> cached_ranks(size); 
-
-    for(int r=0;r<size;r++) {
-      cached_ranks[r]=r;
-    }
-
-    communicator_cached=comm;
-    
-    MPI_Comm_group(communicator_cached, &CachedGroup);
-    
-    MPI_Group_translate_ranks(CachedGroup,size,&cached_ranks[0],WorldGroup, &UserCommunicatorToWorldRanks[0]); 
-    
-    comm_world_peer = UserCommunicatorToWorldRanks[rank];
-    //    std::cout << " MapCommRankToWorldRank  cache miss " <<rank<<"->"<<comm_world_peer<<std::endl;
-    
-    assert(comm_world_peer != MPI_UNDEFINED);
-  }
-
-  assert( (tag & (~0xFFFFL)) ==0); 
-  
-  uint64_t icomm = (uint64_t)comm;
-  int comm_hash = ((icomm>>0 )&0xFFFF)^((icomm>>16)&0xFFFF)
-                ^ ((icomm>>32)&0xFFFF)^((icomm>>48)&0xFFFF);
-  
-  //  hashtag = (comm_hash<<15) | tag;      
-  hashtag = tag;      
-
-};
-
-void Slave::Init(SlaveState * _state,MPI_Comm _squadron,int _universe_rank,int _vertical_rank)
-{
-  squadron=_squadron;
-  universe_rank=_universe_rank;
-  vertical_rank=_vertical_rank;
-  state   =_state;
-  //  std::cout << "state "<<_state<<" comm "<<_squadron<<" universe_rank"<<universe_rank <<std::endl;
-  state->head = state->tail = state->start = 0;
-  base = (uint64_t)MPIoffloadEngine::VerticalShmBufs[0];
-  int rank; MPI_Comm_rank(_squadron,&rank);
-}
-#define PERI_PLUS(A) ( (A+1)%pool )
-int Slave::Event (void) {
-
-  static int tail_last;
-  static int head_last;
-  static int start_last;
-  int ierr;
-  MPI_Status stat;
-  static int i=0;
-
-  ////////////////////////////////////////////////////
-  // Try to advance the start pointers
-  ////////////////////////////////////////////////////
-  int s=state->start;
-  if ( s != state->head ) {
-    switch ( state->Descrs[s].command ) {
-    case COMMAND_ISEND:
-      ierr = MPI_Isend((void *)(state->Descrs[s].buf+base), 
-		       state->Descrs[s].bytes, 
-		       MPI_CHAR,
-		       state->Descrs[s].rank,
-		       state->Descrs[s].tag,
-		       MPIoffloadEngine::communicator_universe,
-		       (MPI_Request *)&state->Descrs[s].request);
-      assert(ierr==0);
-      state->start = PERI_PLUS(s);
-      return 1;
-      break;
-
-    case COMMAND_IRECV:
-      ierr=MPI_Irecv((void *)(state->Descrs[s].buf+base), 
-		     state->Descrs[s].bytes, 
-		     MPI_CHAR,
-		     state->Descrs[s].rank,
-		     state->Descrs[s].tag,
-		     MPIoffloadEngine::communicator_universe,
-		     (MPI_Request *)&state->Descrs[s].request);
-
-      //      std::cout<< " Request is "<<state->Descrs[s].request<<std::endl;
-      //      std::cout<< " Request0 is "<<state->Descrs[0].request<<std::endl;
-      assert(ierr==0);
-      state->start = PERI_PLUS(s);
-      return 1;
-      break;
-
-    case COMMAND_SENDRECV:
-
-      //      fprintf(stderr,"Sendrecv ->%d %d : <-%d %d \n",state->Descrs[s].dest, state->Descrs[s].xtag+i*10,state->Descrs[s].src, state->Descrs[s].rtag+i*10);
-
-      ierr=MPI_Sendrecv((void *)(state->Descrs[s].xbuf+base), state->Descrs[s].bytes, MPI_CHAR, state->Descrs[s].dest, state->Descrs[s].xtag+i*10,
-			(void *)(state->Descrs[s].rbuf+base), state->Descrs[s].bytes, MPI_CHAR, state->Descrs[s].src , state->Descrs[s].rtag+i*10,
-			MPIoffloadEngine::communicator_universe,MPI_STATUS_IGNORE);
-
-      assert(ierr==0);
-
-      //      fprintf(stderr,"Sendrecv done %d %d\n",ierr,i);
-      //      MPI_Barrier(MPIoffloadEngine::HorizontalComm);
-      //      fprintf(stderr,"Barrier\n");
-      i++;
-
-      state->start = PERI_PLUS(s);
-
-      return 1;
-      break;
-
-    case COMMAND_WAITALL:
-
-      for(int t=state->tail;t!=s; t=PERI_PLUS(t) ){
-	if ( state->Descrs[t].command != COMMAND_SENDRECV ) {
-	  MPI_Wait((MPI_Request *)&state->Descrs[t].request,MPI_STATUS_IGNORE);
-	}
-      };
-      s=PERI_PLUS(s);
-      state->start = s;
-      state->tail  = s;
-
-      WakeUpCompute();
-
-      return 1;
-      break;
-
-    default:
-      assert(0);
-      break;
-    }
-  }
-  return 0;
-}
-  //////////////////////////////////////////////////////////////////////////////
-  // External interaction with the queue
-  //////////////////////////////////////////////////////////////////////////////
-  
-void Slave::QueueSendRecv(void *xbuf, void *rbuf, int bytes, int xtag, int rtag, MPI_Comm comm,int dest,int src) 
-{
-  int head =state->head;
-  int next = PERI_PLUS(head);
-  
-  // Set up descriptor
-  int worldrank;
-  int hashtag;
-  MPI_Comm    communicator;
-  MPI_Request request;
-  uint64_t relative;
-  
-  relative = (uint64_t)xbuf - base;
-  state->Descrs[head].xbuf    = relative;
-  
-  relative= (uint64_t)rbuf - base;
-  state->Descrs[head].rbuf    = relative;
-  
-  state->Descrs[head].bytes  = bytes;
-  
-  MPIoffloadEngine::MapCommRankToWorldRank(hashtag,worldrank,xtag,comm,dest);
-  state->Descrs[head].dest   = MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank];
-  state->Descrs[head].xtag    = hashtag;
-  
-  MPIoffloadEngine::MapCommRankToWorldRank(hashtag,worldrank,rtag,comm,src);
-  state->Descrs[head].src    = MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank];
-  state->Descrs[head].rtag    = hashtag;
-  
-  state->Descrs[head].command= COMMAND_SENDRECV;
-  
-  // Block until FIFO has space
-  while( state->tail==next );
-  
-  // Msync on weak order architectures
-  
-  // Advance pointer
-  state->head = next;
-  
-};
-uint64_t Slave::QueueCommand(int command,void *buf, int bytes, int tag, MPI_Comm comm,int commrank) 
-{
-  /////////////////////////////////////////
-  // Spin; if FIFO is full until not full
-  /////////////////////////////////////////
-  int head =state->head;
-  int next = PERI_PLUS(head);
-    
-  // Set up descriptor
-  int worldrank;
-  int hashtag;
-  MPI_Comm    communicator;
-  MPI_Request request;
-  
-  MPIoffloadEngine::MapCommRankToWorldRank(hashtag,worldrank,tag,comm,commrank);
-
-  uint64_t relative= (uint64_t)buf - base;
-  state->Descrs[head].buf    = relative;
-  state->Descrs[head].bytes  = bytes;
-  state->Descrs[head].rank   = MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank];
-  state->Descrs[head].tag    = hashtag;
-  state->Descrs[head].command= command;
-
-  /*  
-  if ( command == COMMAND_ISEND ) { 
-  std::cout << "QueueSend from "<< universe_rank <<" to commrank " << commrank 
-            << " to worldrank " << worldrank <<std::endl;
-  std::cout << " via VerticalRank "<< vertical_rank <<" to universerank " << MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank]<<std::endl;
-  std::cout << " QueueCommand "<<buf<<"["<<bytes<<"]" << std::endl;
-  } 
-  if ( command == COMMAND_IRECV ) { 
-  std::cout << "QueueRecv on "<< universe_rank <<" from commrank " << commrank 
-            << " from worldrank " << worldrank <<std::endl;
-  std::cout << " via VerticalRank "<< vertical_rank <<" from universerank " << MPIoffloadEngine::UniverseRanks[worldrank][vertical_rank]<<std::endl;
-  std::cout << " QueueSend "<<buf<<"["<<bytes<<"]" << std::endl;
-  } 
-  */
-  // Block until FIFO has space
-  while( state->tail==next );
-
-  // Msync on weak order architectures
-  // Advance pointer
-  state->head = next;
-
-  return 0;
-}
-  
-
-///////////////////////////////////////////////////////////////////////////////////////////////////
-// Info that is setup once and indept of cartesian layout
-///////////////////////////////////////////////////////////////////////////////////////////////////
-
-MPI_Comm CartesianCommunicator::communicator_world;
-
-void CartesianCommunicator::Init(int *argc, char ***argv) 
-{
-  int flag;
-  MPI_Initialized(&flag); // needed to coexist with other libs apparently
-  if ( !flag ) {
-    MPI_Init(argc,argv);
-  }
-  communicator_world = MPI_COMM_WORLD;
-  MPI_Comm ShmComm;
-  MPIoffloadEngine::CommunicatorInit (communicator_world,ShmComm,ShmCommBuf);
-}
-void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
-{
-  int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
-  assert(ierr==0);
-}
-int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
-{
-  int rank;
-  int ierr=MPI_Cart_rank  (communicator, &coor[0], &rank);
-  assert(ierr==0);
-  return rank;
-}
-void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
-{
-  coor.resize(_ndimension);
-  int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
-  assert(ierr==0);
-}
-
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
-{ 
-  _ndimension = processors.size();
-  std::vector<int> periodic(_ndimension,1);
-
-  _Nprocessors=1;
-  _processors = processors;
-
-  for(int i=0;i<_ndimension;i++){
-    _Nprocessors*=_processors[i];
-  }
-
-  int Size; 
-  MPI_Comm_size(communicator_world,&Size);
-  assert(Size==_Nprocessors);
-
-  _processor_coor.resize(_ndimension);
-  MPI_Cart_create(communicator_world, _ndimension,&_processors[0],&periodic[0],1,&communicator);
-  MPI_Comm_rank  (communicator,&_processor);
-  MPI_Cart_coords(communicator,_processor,_ndimension,&_processor_coor[0]);
-};
-
-void CartesianCommunicator::GlobalSum(uint32_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(uint64_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(float &f){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(float *f,int N)
-{
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(double &d)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(double *d,int N)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-
-// Basic Halo comms primitive
-void CartesianCommunicator::SendToRecvFrom(void *xmit,
-					   int dest,
-					   void *recv,
-					   int from,
-					   int bytes)
-{
-  std::vector<CommsRequest_t> reqs(0);
-  SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
-  SendToRecvFromComplete(reqs);
-}
-
-void CartesianCommunicator::SendRecvPacket(void *xmit,
-					   void *recv,
-					   int sender,
-					   int receiver,
-					   int bytes)
-{
-  MPI_Status stat;
-  assert(sender != receiver);
-  int tag = sender;
-  if ( _processor == sender ) {
-    MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
-  }
-  if ( _processor == receiver ) { 
-    MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
-  }
-}
-
-// Basic Halo comms primitive
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						void *xmit,
-						int dest,
-						void *recv,
-						int from,
-						int bytes)
-{
-  MPI_Request xrq;
-  MPI_Request rrq;
-  int rank = _processor;
-  int ierr;
-  ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
-  ierr|=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
-  
-  assert(ierr==0);
-
-  list.push_back(xrq);
-  list.push_back(rrq);
-}
-
-void CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						       void *xmit,
-						       int dest,
-						       void *recv,
-						       int from,
-						       int bytes)
-{
-  uint64_t xmit_i = (uint64_t) xmit;
-  uint64_t recv_i = (uint64_t) recv;
-  uint64_t shm    = (uint64_t) ShmCommBuf;
-  // assert xmit and recv lie in shared memory region
-  assert( (xmit_i >= shm) && (xmit_i+bytes <= shm+MAX_MPI_SHM_BYTES) );
-  assert( (recv_i >= shm) && (recv_i+bytes <= shm+MAX_MPI_SHM_BYTES) );
-  assert(from!=_processor);
-  assert(dest!=_processor);
-
-  MPIoffloadEngine::QueueMultiplexedSendRecv(xmit,recv,bytes,_processor,from,communicator,dest,from);
-
-  //MPIoffloadEngine::QueueRoundRobinSendRecv(xmit,recv,bytes,_processor,from,communicator,dest,from);
-
-  //MPIoffloadEngine::QueueMultiplexedSend(xmit,bytes,_processor,communicator,dest);
-  //MPIoffloadEngine::QueueMultiplexedRecv(recv,bytes,from,communicator,from);
-}
-
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list)
-{
-  MPIoffloadEngine::WaitAll();
-  //this->Barrier();
-}
-
-void CartesianCommunicator::StencilBarrier(void) { }
-
-void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
-{
-  int nreq=list.size();
-  std::vector<MPI_Status> status(nreq);
-  int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-  assert(ierr==0);
-}
-
-void CartesianCommunicator::Barrier(void)
-{
-  int ierr = MPI_Barrier(communicator);
-  assert(ierr==0);
-}
-
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
-{
-  int ierr=MPI_Bcast(data,
-		     bytes,
-		     MPI_BYTE,
-		     root,
-		     communicator);
-  assert(ierr==0);
-}
-
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
-{
-  int ierr= MPI_Bcast(data,
-		      bytes,
-		      MPI_BYTE,
-		      root,
-		      communicator_world);
-  assert(ierr==0);
-}
-
-void *CartesianCommunicator::ShmBufferSelf(void) { return ShmCommBuf; }
-
-void *CartesianCommunicator::ShmBuffer(int rank) {
-  return NULL;
-}
-void *CartesianCommunicator::ShmBufferTranslate(int rank,void * local_p) { 
-  return NULL;
-}
-
-
-};
-

lib/communicator/Communicator_mpit.cc

@@ -1,259 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/communicator/Communicator_mpi.cc
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#include <Grid/GridCore.h>
-#include <Grid/GridQCDcore.h>
-#include <Grid/qcd/action/ActionCore.h>
-#include <mpi.h>
-
-namespace Grid {
-
-
-///////////////////////////////////////////////////////////////////////////////////////////////////
-// Info that is setup once and indept of cartesian layout
-///////////////////////////////////////////////////////////////////////////////////////////////////
-MPI_Comm CartesianCommunicator::communicator_world;
-
-// Should error check all MPI calls.
-void CartesianCommunicator::Init(int *argc, char ***argv) {
-  int flag;
-  int provided;
-  MPI_Initialized(&flag); // needed to coexist with other libs apparently
-  if ( !flag ) {
-    MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
-    if ( provided != MPI_THREAD_MULTIPLE ) {
-      QCD::WilsonKernelsStatic::Comms = QCD::WilsonKernelsStatic::CommsThenCompute;
-    }
-  }
-  MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
-  ShmInitGeneric();
-}
-
-void CartesianCommunicator::GlobalSum(uint32_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(uint64_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalXOR(uint32_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT32_T,MPI_BXOR,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalXOR(uint64_t &u){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&u,1,MPI_UINT64_T,MPI_BXOR,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(float &f){
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,&f,1,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(float *f,int N)
-{
-  int ierr=MPI_Allreduce(MPI_IN_PLACE,f,N,MPI_FLOAT,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSum(double &d)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,&d,1,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::GlobalSumVector(double *d,int N)
-{
-  int ierr = MPI_Allreduce(MPI_IN_PLACE,d,N,MPI_DOUBLE,MPI_SUM,communicator);
-  assert(ierr==0);
-}
-void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
-{
-  int ierr=MPI_Cart_shift(communicator,dim,shift,&source,&dest);
-  assert(ierr==0);
-}
-int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
-{
-  int rank;
-  int ierr=MPI_Cart_rank  (communicator, &coor[0], &rank);
-  assert(ierr==0);
-  return rank;
-}
-void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
-{
-  coor.resize(_ndimension);
-  int ierr=MPI_Cart_coords  (communicator, rank, _ndimension,&coor[0]);
-  assert(ierr==0);
-}
-
-// Basic Halo comms primitive
-void CartesianCommunicator::SendToRecvFrom(void *xmit,
-					   int dest,
-					   void *recv,
-					   int from,
-					   int bytes)
-{
-  std::vector<CommsRequest_t> reqs(0);
-  SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
-  SendToRecvFromComplete(reqs);
-}
-
-void CartesianCommunicator::SendRecvPacket(void *xmit,
-					   void *recv,
-					   int sender,
-					   int receiver,
-					   int bytes)
-{
-  MPI_Status stat;
-  assert(sender != receiver);
-  int tag = sender;
-  if ( _processor == sender ) {
-    MPI_Send(xmit, bytes, MPI_CHAR,receiver,tag,communicator);
-  }
-  if ( _processor == receiver ) { 
-    MPI_Recv(recv, bytes, MPI_CHAR,sender,tag,communicator,&stat);
-  }
-}
-
-// Basic Halo comms primitive
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						void *xmit,
-						int dest,
-						void *recv,
-						int from,
-						int bytes)
-{
-  int myrank = _processor;
-  int ierr;
-  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
-    MPI_Request xrq;
-    MPI_Request rrq;
-
-    ierr =MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
-    ierr|=MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
-    
-    assert(ierr==0);
-    list.push_back(xrq);
-    list.push_back(rrq);
-  } else { 
-    // Give the CPU to MPI immediately; can use threads to overlap optionally
-    ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
-		      recv,bytes,MPI_CHAR,from, from,
-		      communicator,MPI_STATUS_IGNORE);
-    assert(ierr==0);
-  }
-}
-void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
-{
-  if ( CommunicatorPolicy == CommunicatorPolicyConcurrent ) { 
-    int nreq=list.size();
-    std::vector<MPI_Status> status(nreq);
-    int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
-    assert(ierr==0);
-  }
-}
-
-void CartesianCommunicator::Barrier(void)
-{
-  int ierr = MPI_Barrier(communicator);
-  assert(ierr==0);
-}
-
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
-{
-  int ierr=MPI_Bcast(data,
-		     bytes,
-		     MPI_BYTE,
-		     root,
-		     communicator);
-  assert(ierr==0);
-}
-  ///////////////////////////////////////////////////////
-  // Should only be used prior to Grid Init finished.
-  // Check for this?
-  ///////////////////////////////////////////////////////
-int CartesianCommunicator::RankWorld(void){ 
-  int r; 
-  MPI_Comm_rank(communicator_world,&r);
-  return r;
-}
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
-{
-  int ierr= MPI_Bcast(data,
-		      bytes,
-		      MPI_BYTE,
-		      root,
-		      communicator_world);
-  assert(ierr==0);
-}
-
-double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-							 void *xmit,
-							 int xmit_to_rank,
-							 void *recv,
-							 int recv_from_rank,
-							 int bytes,int dir)
-{
-  int myrank = _processor;
-  int ierr;
-  assert(dir < communicator_halo.size());
-  
-  //  std::cout << " sending on communicator "<<dir<<" " <<communicator_halo[dir]<<std::endl;
-  // Give the CPU to MPI immediately; can use threads to overlap optionally
-  MPI_Request req[2];
-  MPI_Irecv(recv,bytes,MPI_CHAR,recv_from_rank,recv_from_rank, communicator_halo[dir],&req[1]);
-  MPI_Isend(xmit,bytes,MPI_CHAR,xmit_to_rank  ,myrank        , communicator_halo[dir],&req[0]);
-
-  list.push_back(req[0]);
-  list.push_back(req[1]);
-  return 2.0*bytes;
-}
-void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
-{ 
-  int nreq=waitall.size();
-  MPI_Waitall(nreq, &waitall[0], MPI_STATUSES_IGNORE);
-};
-double CartesianCommunicator::StencilSendToRecvFrom(void *xmit,
-						    int xmit_to_rank,
-						    void *recv,
-						    int recv_from_rank,
-						    int bytes,int dir)
-{
-  int myrank = _processor;
-  int ierr;
-  assert(dir < communicator_halo.size());
-  
-  //  std::cout << " sending on communicator "<<dir<<" " <<communicator_halo[dir]<<std::endl;
-  // Give the CPU to MPI immediately; can use threads to overlap optionally
-  MPI_Request req[2];
-  MPI_Irecv(recv,bytes,MPI_CHAR,recv_from_rank,recv_from_rank, communicator_halo[dir],&req[1]);
-  MPI_Isend(xmit,bytes,MPI_CHAR,xmit_to_rank  ,myrank        , communicator_halo[dir],&req[0]);
-  MPI_Waitall(2, req, MPI_STATUSES_IGNORE);
-  return 2.0*bytes;
-}
-
-
-
-}
-

lib/communicator/Communicator_none.cc

@@ -32,14 +32,22 @@ namespace Grid {
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 // Info that is setup once and indept of cartesian layout
 ///////////////////////////////////////////////////////////////////////////////////////////////////
+Grid_MPI_Comm       CartesianCommunicator::communicator_world;
 
 void CartesianCommunicator::Init(int *argc, char *** arv)
 {
-  ShmInitGeneric();
+  GlobalSharedMemory::Init(communicator_world);
+  GlobalSharedMemory::SharedMemoryAllocate(
+		   GlobalSharedMemory::MAX_MPI_SHM_BYTES,
+		   GlobalSharedMemory::Hugepages);
 }
 
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent) 
+CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent,int &srank) 
-  : CartesianCommunicator(processors) {}
+  : CartesianCommunicator(processors) 
+{
+  srank=0;
+  SetCommunicator(communicator_world);
+}
 
 CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
 {
@@ -54,8 +62,11 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
     assert(_processors[d]==1);
     _processor_coor[d] = 0;
   }
+  SetCommunicator(communicator_world);
 }
 
+CartesianCommunicator::~CartesianCommunicator(){}
+
 void CartesianCommunicator::GlobalSum(float &){}
 void CartesianCommunicator::GlobalSumVector(float *,int N){}
 void CartesianCommunicator::GlobalSum(double &){}
@@ -98,6 +109,14 @@ void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &
 {
   assert(0);
 }
+void CartesianCommunicator::AllToAll(int dim,void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  bcopy(in,out,bytes*words);
+}
+void CartesianCommunicator::AllToAll(void  *in,void *out,uint64_t words,uint64_t bytes)
+{
+  bcopy(in,out,bytes*words);
+}
 
 int  CartesianCommunicator::RankWorld(void){return 0;}
 void CartesianCommunicator::Barrier(void){}
@@ -111,6 +130,36 @@ void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest
   dest=0;
 }
 
+double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
+						     int xmit_to_rank,
+						     void *recv,
+						     int recv_from_rank,
+						     int bytes, int dir)
+{
+  std::vector<CommsRequest_t> list;
+  // Discard the "dir"
+  SendToRecvFromBegin   (list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
+  SendToRecvFromComplete(list);
+  return 2.0*bytes;
+}
+double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
+							 void *xmit,
+							 int xmit_to_rank,
+							 void *recv,
+							 int recv_from_rank,
+							 int bytes, int dir)
+{
+  // Discard the "dir"
+  SendToRecvFromBegin(list,xmit,xmit_to_rank,recv,recv_from_rank,bytes);
+  return 2.0*bytes;
+}
+void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
+{
+  SendToRecvFromComplete(waitall);
+}
+
+void CartesianCommunicator::StencilBarrier(void){};
+
 
 }
 

lib/communicator/Communicator_shmem.cc

@@ -1,355 +0,0 @@
-    /*************************************************************************************
-
-    Grid physics library, www.github.com/paboyle/Grid 
-
-    Source file: ./lib/communicator/Communicator_shmem.cc
-
-    Copyright (C) 2015
-
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-
-    This program is free software; you can redistribute it and/or modify
-    it under the terms of the GNU General Public License as published by
-    the Free Software Foundation; either version 2 of the License, or
-    (at your option) any later version.
-
-    This program is distributed in the hope that it will be useful,
-    but WITHOUT ANY WARRANTY; without even the implied warranty of
-    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
-    GNU General Public License for more details.
-
-    You should have received a copy of the GNU General Public License along
-    with this program; if not, write to the Free Software Foundation, Inc.,
-    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
-
-    See the full license in the file "LICENSE" in the top level distribution directory
-    *************************************************************************************/
-    /*  END LEGAL */
-#include <Grid/Grid.h>
-#include <mpp/shmem.h>
-#include <array>
-
-namespace Grid {
-
-  // Should error check all MPI calls.
-#define SHMEM_VET(addr) 
-
-#define SHMEM_VET_DEBUG(addr) {				\
-  if ( ! shmem_addr_accessible(addr,_processor) ) {\
-    std::fprintf(stderr,"%d Inaccessible shmem address %lx %s %s\n",_processor,addr,__FUNCTION__,#addr); \
-    BACKTRACEFILE();		   \
-  }\
-}
-
-
-///////////////////////////////////////////////////////////////////////////////////////////////////
-// Info that is setup once and indept of cartesian layout
-///////////////////////////////////////////////////////////////////////////////////////////////////
-
-typedef struct HandShake_t { 
-  uint64_t seq_local;
-  uint64_t seq_remote;
-} HandShake;
-
-std::array<long,_SHMEM_REDUCE_SYNC_SIZE> make_psync_init(void) {
-  std::array<long,_SHMEM_REDUCE_SYNC_SIZE> ret;
-  ret.fill(SHMEM_SYNC_VALUE);
-  return ret;
-}
-static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync_init = make_psync_init();
-
-static Vector< HandShake > XConnections;
-static Vector< HandShake > RConnections;
-
-void CartesianCommunicator::Init(int *argc, char ***argv) {
-  shmem_init();
-  XConnections.resize(shmem_n_pes());
-  RConnections.resize(shmem_n_pes());
-  for(int pe =0 ; pe<shmem_n_pes();pe++){
-    XConnections[pe].seq_local = 0;
-    XConnections[pe].seq_remote= 0;
-    RConnections[pe].seq_local = 0;
-    RConnections[pe].seq_remote= 0;
-  }
-  shmem_barrier_all();
-  ShmInitGeneric();
-}
-
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,const CartesianCommunicator &parent) 
-  : CartesianCommunicator(processors) 
-{
-  std::cout << "Attempts to split SHMEM communicators will fail " <<std::endl;
-}
-CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
-{
-  _ndimension = processors.size();
-  std::vector<int> periodic(_ndimension,1);
-
-  _Nprocessors=1;
-  _processors = processors;
-  _processor_coor.resize(_ndimension);
-
-  _processor = shmem_my_pe();
-  
-  Lexicographic::CoorFromIndex(_processor_coor,_processor,_processors);
-
-  for(int i=0;i<_ndimension;i++){
-    _Nprocessors*=_processors[i];
-  }
-
-  int Size = shmem_n_pes(); 
-
-
-  assert(Size==_Nprocessors);
-}
-
-void CartesianCommunicator::GlobalSum(uint32_t &u){
-  static long long source ;
-  static long long dest   ;
-  static long long llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
-  static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync =  psync_init;
-
-  //  int nreduce=1;
-  //  int pestart=0;
-  //  int logStride=0;
-
-  source = u;
-  dest   = 0;
-  shmem_longlong_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
-  shmem_barrier_all(); // necessary?
-  u = dest;
-}
-void CartesianCommunicator::GlobalSum(uint64_t &u){
-  static long long source ;
-  static long long dest   ;
-  static long long llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
-  static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync =  psync_init;
-
-  //  int nreduce=1;
-  //  int pestart=0;
-  //  int logStride=0;
-
-  source = u;
-  dest   = 0;
-  shmem_longlong_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
-  shmem_barrier_all(); // necessary?
-  u = dest;
-}
-void CartesianCommunicator::GlobalSum(float &f){
-  static float source ;
-  static float dest   ;
-  static float llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
-  static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync =  psync_init;
-
-  source = f;
-  dest   =0.0;
-  shmem_float_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
-  shmem_barrier_all();
-  f = dest;
-}
-void CartesianCommunicator::GlobalSumVector(float *f,int N)
-{
-  static float source ;
-  static float dest   = 0 ;
-  static float llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
-  static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync =  psync_init;
-
-  if ( shmem_addr_accessible(f,_processor)  ){
-    shmem_float_sum_to_all(f,f,N,0,0,_Nprocessors,llwrk,psync.data());
-    shmem_barrier_all();
-    return;
-  }
-
-  for(int i=0;i<N;i++){
-    dest   =0.0;
-    source = f[i];
-    shmem_float_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
-    shmem_barrier_all();
-    f[i] = dest;
-  }
-}
-void CartesianCommunicator::GlobalSum(double &d)
-{
-  static double source;
-  static double dest  ;
-  static double llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
-  static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync =  psync_init;
-
-  source = d;
-  dest   = 0;
-  shmem_double_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
-  shmem_barrier_all();
-  d = dest;
-}
-void CartesianCommunicator::GlobalSumVector(double *d,int N)
-{
-  static double source ;
-  static double dest   ;
-  static double llwrk[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];
-  static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync =  psync_init;
-
-
-  if ( shmem_addr_accessible(d,_processor)  ){
-    shmem_double_sum_to_all(d,d,N,0,0,_Nprocessors,llwrk,psync.data());
-    shmem_barrier_all();
-    return;
-  }
-
-  for(int i=0;i<N;i++){
-    source = d[i];
-    dest   =0.0;
-    shmem_double_sum_to_all(&dest,&source,1,0,0,_Nprocessors,llwrk,psync.data());
-    shmem_barrier_all();
-    d[i] = dest;
-  }
-}
-void CartesianCommunicator::ShiftedRanks(int dim,int shift,int &source,int &dest)
-{
-  std::vector<int> coor = _processor_coor;
-
-  assert(std::abs(shift) <_processors[dim]);
-
-  coor[dim] = (_processor_coor[dim] + shift + _processors[dim])%_processors[dim];
-  Lexicographic::IndexFromCoor(coor,source,_processors);
-
-  coor[dim] = (_processor_coor[dim] - shift + _processors[dim])%_processors[dim];
-  Lexicographic::IndexFromCoor(coor,dest,_processors);
-
-}
-int CartesianCommunicator::RankFromProcessorCoor(std::vector<int> &coor)
-{
-  int rank;
-  Lexicographic::IndexFromCoor(coor,rank,_processors);
-  return rank;
-}
-void  CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &coor)
-{
-  Lexicographic::CoorFromIndex(coor,rank,_processors);
-}
-
-// Basic Halo comms primitive
-void CartesianCommunicator::SendToRecvFrom(void *xmit,
-					   int dest,
-					   void *recv,
-					   int from,
-					   int bytes)
-{
-  SHMEM_VET(xmit);
-  SHMEM_VET(recv);
-  std::vector<CommsRequest_t> reqs(0);
-  SendToRecvFromBegin(reqs,xmit,dest,recv,from,bytes);
-  SendToRecvFromComplete(reqs);
-}
-
-void CartesianCommunicator::SendRecvPacket(void *xmit,
-					   void *recv,
-					   int sender,
-					   int receiver,
-					   int bytes)
-{
-  static uint64_t seq;
-
-  assert(recv!=xmit);
-  volatile HandShake *RecvSeq = (volatile HandShake *) & RConnections[sender];
-  volatile HandShake *SendSeq = (volatile HandShake *) & XConnections[receiver];
-
-  if ( _processor == sender ) {
-
-    // Check he has posted a receive
-    while(SendSeq->seq_remote == SendSeq->seq_local);
-
-    // Advance our send count
-    seq = ++(SendSeq->seq_local);
-    
-    // Send this packet 
-    SHMEM_VET(recv);
-    shmem_putmem(recv,xmit,bytes,receiver);
-    shmem_fence();
-
-    //Notify him we're done
-    shmem_putmem((void *)&(RecvSeq->seq_remote),&seq,sizeof(seq),receiver);
-    shmem_fence();
-  }
-  if ( _processor == receiver ) {
-
-    // Post a receive
-    seq = ++(RecvSeq->seq_local);
-    shmem_putmem((void *)&(SendSeq->seq_remote),&seq,sizeof(seq),sender);
-
-    // Now wait until he has advanced our reception counter
-    while(RecvSeq->seq_remote != RecvSeq->seq_local);
-
-  }
-}
-
-// Basic Halo comms primitive
-void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
-						void *xmit,
-						int dest,
-						void *recv,
-						int from,
-						int bytes)
-{
-  SHMEM_VET(xmit);
-  SHMEM_VET(recv);
-  //  shmem_putmem_nb(recv,xmit,bytes,dest,NULL);
-  shmem_putmem(recv,xmit,bytes,dest);
-
-  if ( CommunicatorPolicy == CommunicatorPolicySequential ) shmem_barrier_all(); 
-}
-void CartesianCommunicator::SendToRecvFromComplete(std::vector<CommsRequest_t> &list)
-{
-  //  shmem_quiet();      // I'm done
-  if( CommunicatorPolicy == CommunicatorPolicyConcurrent ) shmem_barrier_all();// He's done too
-}
-void CartesianCommunicator::Barrier(void)
-{
-  shmem_barrier_all();
-}
-void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
-{
-  static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync =  psync_init;
-  static uint32_t word;
-  uint32_t *array = (uint32_t *) data;
-  assert( (bytes % 4)==0);
-  int words = bytes/4;
-
-  if ( shmem_addr_accessible(data,_processor)  ){
-    shmem_broadcast32(data,data,words,root,0,0,shmem_n_pes(),psync.data());
-    return;
-  }
-
-  for(int w=0;w<words;w++){
-    word = array[w];
-    shmem_broadcast32((void *)&word,(void *)&word,1,root,0,0,shmem_n_pes(),psync.data());
-    if ( shmem_my_pe() != root ) {
-      array[w] = word;
-    }
-    shmem_barrier_all();
-  }
-
-}
-void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
-{
-  static std::array<long,_SHMEM_REDUCE_SYNC_SIZE> psync =  psync_init;
-  static uint32_t word;
-  uint32_t *array = (uint32_t *) data;
-  assert( (bytes % 4)==0);
-  int words = bytes/4;
-
-  for(int w=0;w<words;w++){
-    word = array[w];
-    shmem_broadcast32((void *)&word,(void *)&word,1,root,0,0,shmem_n_pes(),psync.data());
-    if ( shmem_my_pe() != root ) {
-      array[w]= word;
-    }
-    shmem_barrier_all();
-  }
-}
-  
-int CartesianCommunicator::RankWorld(void){ 
-  return shmem_my_pe();
-}
-
-}
-

lib/communicator/SharedMemory.cc

@@ -0,0 +1,92 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/SharedMemory.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/GridCore.h>
+
+namespace Grid { 
+
+// static data
+
+uint64_t            GlobalSharedMemory::MAX_MPI_SHM_BYTES   = 1024LL*1024LL*1024LL; 
+int                 GlobalSharedMemory::Hugepages = 0;
+int                 GlobalSharedMemory::_ShmSetup;
+int                 GlobalSharedMemory::_ShmAlloc;
+uint64_t            GlobalSharedMemory::_ShmAllocBytes;
+
+std::vector<void *> GlobalSharedMemory::WorldShmCommBufs;
+
+Grid_MPI_Comm       GlobalSharedMemory::WorldShmComm;
+int                 GlobalSharedMemory::WorldShmRank;
+int                 GlobalSharedMemory::WorldShmSize;
+std::vector<int>    GlobalSharedMemory::WorldShmRanks;
+
+Grid_MPI_Comm       GlobalSharedMemory::WorldComm;
+int                 GlobalSharedMemory::WorldSize;
+int                 GlobalSharedMemory::WorldRank;
+
+int                 GlobalSharedMemory::WorldNodes;
+int                 GlobalSharedMemory::WorldNode;
+
+void GlobalSharedMemory::SharedMemoryFree(void)
+{
+  assert(_ShmAlloc);
+  assert(_ShmAllocBytes>0);
+  for(int r=0;r<WorldShmSize;r++){
+    munmap(WorldShmCommBufs[r],_ShmAllocBytes);
+  }
+  _ShmAlloc = 0;
+  _ShmAllocBytes = 0;
+}
+/////////////////////////////////
+// Alloc, free shmem region
+/////////////////////////////////
+void *SharedMemory::ShmBufferMalloc(size_t bytes){
+  //  bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
+  void *ptr = (void *)heap_top;
+  heap_top  += bytes;
+  heap_bytes+= bytes;
+  if (heap_bytes >= heap_size) {
+    std::cout<< " ShmBufferMalloc exceeded shared heap size -- try increasing with --shm <MB> flag" <<std::endl;
+    std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
+    std::cout<< " Current value is " << (heap_size/(1024*1024)) <<std::endl;
+    assert(heap_bytes<heap_size);
+  }
+  return ptr;
+}
+void SharedMemory::ShmBufferFreeAll(void) { 
+  heap_top  =(size_t)ShmBufferSelf();
+  heap_bytes=0;
+}
+void *SharedMemory::ShmBufferSelf(void)
+{
+  return ShmCommBufs[ShmRank];
+}
+
+
+
+}

lib/communicator/SharedMemory.h

@@ -0,0 +1,164 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/SharedMemory.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+
+// TODO
+// 1) move includes into SharedMemory.cc
+//
+// 2) split shared memory into a) optimal communicator creation from comm world
+// 
+//                             b) shared memory buffers container
+//                                -- static globally shared; init once
+//                                -- per instance set of buffers.
+//                                   
+
+#pragma once 
+
+#include <Grid/GridCore.h>
+
+#if defined (GRID_COMMS_MPI3) 
+#include <mpi.h>
+#endif 
+#include <semaphore.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <limits.h>
+#include <sys/types.h>
+#include <sys/ipc.h>
+#include <sys/shm.h>
+#include <sys/mman.h>
+#include <zlib.h>
+#ifdef HAVE_NUMAIF_H
+#include <numaif.h>
+#endif
+
+namespace Grid {
+
+#if defined (GRID_COMMS_MPI3) 
+  typedef MPI_Comm    Grid_MPI_Comm;
+  typedef MPI_Request CommsRequest_t;
+#else 
+  typedef int CommsRequest_t;
+  typedef int Grid_MPI_Comm;
+#endif
+
+class GlobalSharedMemory {
+ private:
+  static const int     MAXLOG2RANKSPERNODE = 16;            
+
+  // Init once lock on the buffer allocation
+  static int      _ShmSetup;
+  static int      _ShmAlloc;
+  static uint64_t _ShmAllocBytes;
+
+ public:
+  static int      ShmSetup(void)      { return _ShmSetup; }
+  static int      ShmAlloc(void)      { return _ShmAlloc; }
+  static uint64_t ShmAllocBytes(void) { return _ShmAllocBytes; }
+  static uint64_t      MAX_MPI_SHM_BYTES;
+  static int           Hugepages;
+
+  static std::vector<void *> WorldShmCommBufs;
+
+  static Grid_MPI_Comm WorldComm;
+  static int           WorldRank;
+  static int           WorldSize;
+
+  static Grid_MPI_Comm WorldShmComm;
+  static int           WorldShmRank;
+  static int           WorldShmSize;
+
+  static int           WorldNodes;
+  static int           WorldNode;
+
+  static std::vector<int>  WorldShmRanks;
+
+  //////////////////////////////////////////////////////////////////////////////////////
+  // Create an optimal reordered communicator that makes MPI_Cart_create get it right
+  //////////////////////////////////////////////////////////////////////////////////////
+  static void Init(Grid_MPI_Comm comm); // Typically MPI_COMM_WORLD
+  static void OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm);  // Turns MPI_COMM_WORLD into right layout for Cartesian
+  ///////////////////////////////////////////////////
+  // Provide shared memory facilities off comm world
+  ///////////////////////////////////////////////////
+  static void SharedMemoryAllocate(uint64_t bytes, int flags);
+  static void SharedMemoryFree(void);
+
+};
+
+//////////////////////////////
+// one per communicator
+//////////////////////////////
+class SharedMemory 
+{
+ private:
+  static const int     MAXLOG2RANKSPERNODE = 16;            
+
+  size_t heap_top;
+  size_t heap_bytes;
+  size_t heap_size;
+
+ protected:
+
+  Grid_MPI_Comm    ShmComm; // for barriers
+  int    ShmRank; 
+  int    ShmSize;
+  std::vector<void *> ShmCommBufs;
+  std::vector<int>    ShmRanks;// Mapping comm ranks to Shm ranks
+
+ public:
+  SharedMemory() {};
+  ///////////////////////////////////////////////////////////////////////////////////////
+  // set the buffers & sizes
+  ///////////////////////////////////////////////////////////////////////////////////////
+  void SetCommunicator(Grid_MPI_Comm comm);
+
+  ////////////////////////////////////////////////////////////////////////
+  // For this instance ; disjoint buffer sets between splits if split grid
+  ////////////////////////////////////////////////////////////////////////
+  void ShmBarrier(void); 
+
+  ///////////////////////////////////////////////////
+  // Call on any instance
+  ///////////////////////////////////////////////////
+  void SharedMemoryTest(void);
+  void *ShmBufferSelf(void);
+  void *ShmBuffer    (int rank);
+  void *ShmBufferTranslate(int rank,void * local_p);
+  void *ShmBufferMalloc(size_t bytes);
+  void  ShmBufferFreeAll(void) ;
+  
+  //////////////////////////////////////////////////////////////////////////
+  // Make info on Nodes & ranks and Shared memory available
+  //////////////////////////////////////////////////////////////////////////
+  int NodeCount(void) { return GlobalSharedMemory::WorldNodes;};
+  int RankCount(void) { return GlobalSharedMemory::WorldSize;};
+
+};
+
+}

lib/communicator/SharedMemoryMPI.cc

@@ -0,0 +1,395 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/SharedMemory.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/GridCore.h>
+
+namespace Grid { 
+
+/*Construct from an MPI communicator*/
+void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
+{
+  assert(_ShmSetup==0);
+  WorldComm = comm;
+  MPI_Comm_rank(WorldComm,&WorldRank);
+  MPI_Comm_size(WorldComm,&WorldSize);
+  // WorldComm, WorldSize, WorldRank
+
+  /////////////////////////////////////////////////////////////////////
+  // Split into groups that can share memory
+  /////////////////////////////////////////////////////////////////////
+  MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&WorldShmComm);
+  MPI_Comm_rank(WorldShmComm     ,&WorldShmRank);
+  MPI_Comm_size(WorldShmComm     ,&WorldShmSize);
+  // WorldShmComm, WorldShmSize, WorldShmRank
+
+  // WorldNodes
+  WorldNodes = WorldSize/WorldShmSize;
+  assert( (WorldNodes * WorldShmSize) == WorldSize );
+
+  // FIXME: Check all WorldShmSize are the same ?
+
+  /////////////////////////////////////////////////////////////////////
+  // find world ranks in our SHM group (i.e. which ranks are on our node)
+  /////////////////////////////////////////////////////////////////////
+  MPI_Group WorldGroup, ShmGroup;
+  MPI_Comm_group (WorldComm, &WorldGroup); 
+  MPI_Comm_group (WorldShmComm, &ShmGroup);
+
+  std::vector<int> world_ranks(WorldSize);   for(int r=0;r<WorldSize;r++) world_ranks[r]=r;
+
+  WorldShmRanks.resize(WorldSize); 
+  MPI_Group_translate_ranks (WorldGroup,WorldSize,&world_ranks[0],ShmGroup, &WorldShmRanks[0]); 
+
+  ///////////////////////////////////////////////////////////////////
+  // Identify who is in my group and nominate the leader
+  ///////////////////////////////////////////////////////////////////
+  int g=0;
+  std::vector<int> MyGroup;
+  MyGroup.resize(WorldShmSize);
+  for(int rank=0;rank<WorldSize;rank++){
+    if(WorldShmRanks[rank]!=MPI_UNDEFINED){
+      assert(g<WorldShmSize);
+      MyGroup[g++] = rank;
+    }
+  }
+  
+  std::sort(MyGroup.begin(),MyGroup.end(),std::less<int>());
+  int myleader = MyGroup[0];
+  
+  std::vector<int> leaders_1hot(WorldSize,0);
+  std::vector<int> leaders_group(WorldNodes,0);
+  leaders_1hot [ myleader ] = 1;
+    
+  ///////////////////////////////////////////////////////////////////
+  // global sum leaders over comm world
+  ///////////////////////////////////////////////////////////////////
+  int ierr=MPI_Allreduce(MPI_IN_PLACE,&leaders_1hot[0],WorldSize,MPI_INT,MPI_SUM,WorldComm);
+  assert(ierr==0);
+
+  ///////////////////////////////////////////////////////////////////
+  // find the group leaders world rank
+  ///////////////////////////////////////////////////////////////////
+  int group=0;
+  for(int l=0;l<WorldSize;l++){
+    if(leaders_1hot[l]){
+      leaders_group[group++] = l;
+    }
+  }
+
+  ///////////////////////////////////////////////////////////////////
+  // Identify the node of the group in which I (and my leader) live
+  ///////////////////////////////////////////////////////////////////
+  WorldNode=-1;
+  for(int g=0;g<WorldNodes;g++){
+    if (myleader == leaders_group[g]){
+      WorldNode=g;
+    }
+  }
+  assert(WorldNode!=-1);
+  _ShmSetup=1;
+}
+
+void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+{
+  ////////////////////////////////////////////////////////////////
+  // Assert power of two shm_size.
+  ////////////////////////////////////////////////////////////////
+  int log2size = -1;
+  for(int i=0;i<=MAXLOG2RANKSPERNODE;i++){  
+    if ( (0x1<<i) == WorldShmSize ) {
+      log2size = i;
+      break;
+    }
+  }
+  assert(log2size != -1);
+
+  ////////////////////////////////////////////////////////////////
+  // Identify subblock of ranks on node spreading across dims
+  // in a maximally symmetrical way
+  ////////////////////////////////////////////////////////////////
+  int ndimension              = processors.size();
+  std::vector<int> processor_coor(ndimension);
+  std::vector<int> WorldDims = processors;   std::vector<int> ShmDims  (ndimension,1);  std::vector<int> NodeDims (ndimension);
+  std::vector<int> ShmCoor  (ndimension);    std::vector<int> NodeCoor (ndimension);    std::vector<int> WorldCoor(ndimension);
+  int dim = 0;
+  for(int l2=0;l2<log2size;l2++){
+    while ( (WorldDims[dim] / ShmDims[dim]) <= 1 ) dim=(dim+1)%ndimension;
+    ShmDims[dim]*=2;
+    dim=(dim+1)%ndimension;
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Establish torus of processes and nodes with sub-blockings
+  ////////////////////////////////////////////////////////////////
+  for(int d=0;d<ndimension;d++){
+    NodeDims[d] = WorldDims[d]/ShmDims[d];
+  }
+
+  ////////////////////////////////////////////////////////////////
+  // Check processor counts match
+  ////////////////////////////////////////////////////////////////
+  int Nprocessors=1;
+  for(int i=0;i<ndimension;i++){
+    Nprocessors*=processors[i];
+  }
+  assert(WorldSize==Nprocessors);
+
+  ////////////////////////////////////////////////////////////////
+  // Establish mapping between lexico physics coord and WorldRank
+  ////////////////////////////////////////////////////////////////
+  int rank;
+
+  Lexicographic::CoorFromIndexReversed(NodeCoor,WorldNode   ,NodeDims);
+  Lexicographic::CoorFromIndexReversed(ShmCoor ,WorldShmRank,ShmDims);
+  for(int d=0;d<ndimension;d++) WorldCoor[d] = NodeCoor[d]*ShmDims[d]+ShmCoor[d];
+  Lexicographic::IndexFromCoorReversed(WorldCoor,rank,WorldDims);
+
+  /////////////////////////////////////////////////////////////////
+  // Build the new communicator
+  /////////////////////////////////////////////////////////////////
+  int ierr= MPI_Comm_split(WorldComm,0,rank,&optimal_comm);
+  assert(ierr==0);
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////
+// Hugetlbfs mapping intended
+////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef GRID_MPI3_SHMMMAP
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0);
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  // allocate the shared windows for our group
+  //////////////////////////////////////////////////////////////////////////////////////////////////////////
+  MPI_Barrier(WorldShmComm);
+  WorldShmCommBufs.resize(WorldShmSize);
+  
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // Hugetlbf and others map filesystems as mappable huge pages
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  char shm_name [NAME_MAX];
+  for(int r=0;r<WorldShmSize;r++){
+    
+    sprintf(shm_name,GRID_SHM_PATH "/Grid_mpi3_shm_%d_%d",WorldNode,r);
+    int fd=open(shm_name,O_RDWR|O_CREAT,0666);
+    if ( fd == -1) { 
+      printf("open %s failed\n",shm_name);
+      perror("open hugetlbfs");
+      exit(0);
+    }
+    int mmap_flag = MAP_SHARED ;
+#ifdef MAP_POPULATE    
+    mmap_flag|=MAP_POPULATE;
+#endif
+#ifdef MAP_HUGETLB
+    if ( flags ) mmap_flag |= MAP_HUGETLB;
+#endif
+    void *ptr = (void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag,fd, 0); 
+    if ( ptr == (void *)MAP_FAILED ) {    
+      printf("mmap %s failed\n",shm_name);
+      perror("failed mmap");      assert(0);    
+    }
+    assert(((uint64_t)ptr&0x3F)==0);
+    close(fd);
+    WorldShmCommBufs[r] =ptr;
+  }
+  _ShmAlloc=1;
+  _ShmAllocBytes  = bytes;
+};
+#endif // MMAP
+
+#ifdef GRID_MPI3_SHMOPEN
+////////////////////////////////////////////////////////////////////////////////////////////
+// POSIX SHMOPEN ; as far as I know Linux does not allow EXPLICIT HugePages with this case
+// tmpfs (Larry Meadows says) does not support explicit huge page, and this is used for 
+// the posix shm virtual file system
+////////////////////////////////////////////////////////////////////////////////////////////
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{ 
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0); 
+  MPI_Barrier(WorldShmComm);
+  WorldShmCommBufs.resize(WorldShmSize);
+
+  char shm_name [NAME_MAX];
+  if ( WorldShmRank == 0 ) {
+    for(int r=0;r<WorldShmSize;r++){
+	
+      size_t size = bytes;
+      
+      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldNode,r);
+      
+      shm_unlink(shm_name);
+      int fd=shm_open(shm_name,O_RDWR|O_CREAT,0666);
+      if ( fd < 0 ) {	perror("failed shm_open");	assert(0);      }
+      ftruncate(fd, size);
+	
+      int mmap_flag = MAP_SHARED;
+#ifdef MAP_POPULATE 
+      mmap_flag |= MAP_POPULATE;
+#endif
+#ifdef MAP_HUGETLB
+      if (flags) mmap_flag |= MAP_HUGETLB;
+#endif
+      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
+      
+      if ( ptr == (void * )MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
+      assert(((uint64_t)ptr&0x3F)==0);
+      
+      WorldShmCommBufs[r] =ptr;
+      close(fd);
+    }
+  }
+
+  MPI_Barrier(WorldShmComm);
+  
+  if ( WorldShmRank != 0 ) { 
+    for(int r=0;r<WorldShmSize;r++){
+
+      size_t size = bytes ;
+      
+      sprintf(shm_name,"/Grid_mpi3_shm_%d_%d",WorldNode,r);
+      
+      int fd=shm_open(shm_name,O_RDWR,0666);
+      if ( fd<0 ) {	perror("failed shm_open");	assert(0);      }
+      
+      void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+      if ( ptr == MAP_FAILED ) {       perror("failed mmap");      assert(0);    }
+      assert(((uint64_t)ptr&0x3F)==0);
+      WorldShmCommBufs[r] =ptr;
+
+      close(fd);
+    }
+  }
+  _ShmAlloc=1;
+  _ShmAllocBytes = bytes;
+}
+#endif
+
+  ////////////////////////////////////////////////////////
+  // Global shared functionality finished
+  // Now move to per communicator functionality
+  ////////////////////////////////////////////////////////
+void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
+{
+  int rank, size;
+  MPI_Comm_rank(comm,&rank);
+  MPI_Comm_size(comm,&size);
+  ShmRanks.resize(size);
+
+  /////////////////////////////////////////////////////////////////////
+  // Split into groups that can share memory
+  /////////////////////////////////////////////////////////////////////
+  MPI_Comm_split_type(comm, MPI_COMM_TYPE_SHARED, 0, MPI_INFO_NULL,&ShmComm);
+  MPI_Comm_rank(ShmComm     ,&ShmRank);
+  MPI_Comm_size(ShmComm     ,&ShmSize);
+  ShmCommBufs.resize(ShmSize);
+
+  //////////////////////////////////////////////////////////////////////
+  // Map ShmRank to WorldShmRank and use the right buffer
+  //////////////////////////////////////////////////////////////////////
+  assert (GlobalSharedMemory::ShmAlloc()==1);
+  heap_size = GlobalSharedMemory::ShmAllocBytes();
+  for(int r=0;r<ShmSize;r++){
+
+    uint32_t sr = (r==ShmRank) ? GlobalSharedMemory::WorldRank : 0 ;
+
+    MPI_Allreduce(MPI_IN_PLACE,&sr,1,MPI_UINT32_T,MPI_SUM,comm);
+
+    ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[sr];
+  }
+  ShmBufferFreeAll();
+
+  /////////////////////////////////////////////////////////////////////
+  // find comm ranks in our SHM group (i.e. which ranks are on our node)
+  /////////////////////////////////////////////////////////////////////
+  MPI_Group FullGroup, ShmGroup;
+  MPI_Comm_group (comm   , &FullGroup); 
+  MPI_Comm_group (ShmComm, &ShmGroup);
+
+  std::vector<int> ranks(size);   for(int r=0;r<size;r++) ranks[r]=r;
+  MPI_Group_translate_ranks (FullGroup,size,&ranks[0],ShmGroup, &ShmRanks[0]); 
+}
+//////////////////////////////////////////////////////////////////
+// On node barrier
+//////////////////////////////////////////////////////////////////
+void SharedMemory::ShmBarrier(void)
+{
+  MPI_Barrier  (ShmComm);
+}
+//////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Test the shared memory is working
+//////////////////////////////////////////////////////////////////////////////////////////////////////////
+void SharedMemory::SharedMemoryTest(void)
+{
+  ShmBarrier();
+  if ( ShmRank == 0 ) {
+    for(int r=0;r<ShmSize;r++){
+      uint64_t * check = (uint64_t *) ShmCommBufs[r];
+      check[0] = GlobalSharedMemory::WorldNode;
+      check[1] = r;
+      check[2] = 0x5A5A5A;
+    }
+  }
+  ShmBarrier();
+  for(int r=0;r<ShmSize;r++){
+    uint64_t * check = (uint64_t *) ShmCommBufs[r];
+    
+    assert(check[0]==GlobalSharedMemory::WorldNode);
+    assert(check[1]==r);
+    assert(check[2]==0x5A5A5A);
+    
+  }
+  ShmBarrier();
+}
+
+void *SharedMemory::ShmBuffer(int rank)
+{
+  int gpeer = ShmRanks[rank];
+  if (gpeer == MPI_UNDEFINED){
+    return NULL;
+  } else { 
+    return ShmCommBufs[gpeer];
+  }
+}
+void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
+{
+  static int count =0;
+  int gpeer = ShmRanks[rank];
+  assert(gpeer!=ShmRank); // never send to self
+  if (gpeer == MPI_UNDEFINED){
+    return NULL;
+  } else { 
+    uint64_t offset = (uint64_t)local_p - (uint64_t)ShmCommBufs[ShmRank];
+    uint64_t remote = (uint64_t)ShmCommBufs[gpeer]+offset;
+    return (void *) remote;
+  }
+}
+
+}

lib/communicator/SharedMemoryNone.cc

@@ -0,0 +1,126 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/communicator/SharedMemory.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+*************************************************************************************/
+/*  END LEGAL */
+
+#include <Grid/GridCore.h>
+
+namespace Grid { 
+
+/*Construct from an MPI communicator*/
+void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
+{
+  assert(_ShmSetup==0);
+  WorldComm = 0;
+  WorldRank = 0;
+  WorldSize = 1;
+  WorldShmComm = 0 ;
+  WorldShmRank = 0 ;
+  WorldShmSize = 1 ;
+  WorldNodes   = 1 ;
+  WorldNode    = 0 ;
+  WorldShmRanks.resize(WorldSize); WorldShmRanks[0] = 0;
+  WorldShmCommBufs.resize(1);
+  _ShmSetup=1;
+}
+
+void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,Grid_MPI_Comm & optimal_comm)
+{
+  optimal_comm = WorldComm;
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////
+// Hugetlbfs mapping intended, use anonymous mmap
+////////////////////////////////////////////////////////////////////////////////////////////
+void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
+{
+  void * ShmCommBuf ; 
+  assert(_ShmSetup==1);
+  assert(_ShmAlloc==0);
+  int mmap_flag =0;
+#ifdef MAP_ANONYMOUS
+  mmap_flag = mmap_flag| MAP_SHARED | MAP_ANONYMOUS;
+#endif
+#ifdef MAP_ANON
+  mmap_flag = mmap_flag| MAP_SHARED | MAP_ANON;
+#endif
+#ifdef MAP_HUGETLB
+  if ( flags ) mmap_flag |= MAP_HUGETLB;
+#endif
+  ShmCommBuf =(void *) mmap(NULL, bytes, PROT_READ | PROT_WRITE, mmap_flag, -1, 0); 
+  if (ShmCommBuf == (void *)MAP_FAILED) {
+    perror("mmap failed ");
+    exit(EXIT_FAILURE);  
+  }
+#ifdef MADV_HUGEPAGE
+  if (!Hugepages ) madvise(ShmCommBuf,bytes,MADV_HUGEPAGE);
+#endif
+  bzero(ShmCommBuf,bytes);
+  WorldShmCommBufs[0] = ShmCommBuf;
+  _ShmAllocBytes=bytes;
+  _ShmAlloc=1;
+};
+
+  ////////////////////////////////////////////////////////
+  // Global shared functionality finished
+  // Now move to per communicator functionality
+  ////////////////////////////////////////////////////////
+void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
+{
+  assert(GlobalSharedMemory::ShmAlloc()==1);
+  ShmRanks.resize(1);
+  ShmCommBufs.resize(1);
+  ShmRanks[0] = 0;
+  ShmRank     = 0;
+  ShmSize     = 1;
+  //////////////////////////////////////////////////////////////////////
+  // Map ShmRank to WorldShmRank and use the right buffer
+  //////////////////////////////////////////////////////////////////////
+  ShmCommBufs[0] = GlobalSharedMemory::WorldShmCommBufs[0];
+  heap_size      = GlobalSharedMemory::ShmAllocBytes();
+  ShmBufferFreeAll();
+  return;
+}
+//////////////////////////////////////////////////////////////////
+// On node barrier
+//////////////////////////////////////////////////////////////////
+void SharedMemory::ShmBarrier(void){ return ; }
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Test the shared memory is working
+//////////////////////////////////////////////////////////////////////////////////////////////////////////
+void SharedMemory::SharedMemoryTest(void) { return; }
+
+void *SharedMemory::ShmBuffer(int rank)
+{
+  return NULL;
+}
+void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
+{
+  return NULL;
+}
+
+}

lib/lattice/Lattice_rng.h

@@ -77,9 +77,6 @@ namespace Grid {
 
   
 // merge of April 11 2017
-//<<<<<<< HEAD
-
-
   // this function is necessary for the LS vectorised field
   inline int RNGfillable_general(GridBase *coarse,GridBase *fine)
   {
@@ -91,7 +88,6 @@ namespace Grid {
     // all further divisions are local
     for(int d=0;d<lowerdims;d++) assert(fine->_processors[d]==1);
     for(int d=0;d<rngdims;d++) assert(coarse->_processors[d] == fine->_processors[d+lowerdims]);
-    
 
     // then divide the number of local sites
     // check that the total number of sims agree, meanse the iSites are the same
@@ -102,27 +98,6 @@ namespace Grid {
 
     return fine->lSites() / coarse->lSites();
   }
-
-  /*
-  // Wrap seed_seq to give common interface with random_device
-  class fixedSeed {
-  public:
-    typedef std::seed_seq::result_type result_type;
-    std::seed_seq src;
-    
-    fixedSeed(const std::vector<int> &seeds) : src(seeds.begin(),seeds.end()) {};
-
-    result_type operator () (void){
-      std::vector<result_type> list(1);
-      src.generate(list.begin(),list.end());
-      return list[0];
-    }
-
-  };
-
-=======
->>>>>>> develop
-  */
   
   // real scalars are one component
   template<class scalar,class distribution,class generator> 
@@ -171,7 +146,7 @@ namespace Grid {
     // support for parallel init
     ///////////////////////
 #ifdef RNG_FAST_DISCARD
-    static void Skip(RngEngine &eng)
+    static void Skip(RngEngine &eng,uint64_t site)
     {
       /////////////////////////////////////////////////////////////////////////////////////
       // Skip by 2^40 elements between successive lattice sites
@@ -184,8 +159,11 @@ namespace Grid {
       // and margin of safety is orders of magnitude.
       // We could hack Sitmo to skip in the higher order words of state if necessary
       /////////////////////////////////////////////////////////////////////////////////////
-      uint64_t skip = 0x1; skip = skip<<40;
+      //      uint64_t skip = site+1;  //   Old init Skipped then drew.  Checked compat with faster init
+      uint64_t skip = site;
+      skip = skip<<40;
       eng.discard(skip);
+      //      std::cout << " Engine  " <<site << " state " <<eng<<std::endl;
     } 
 #endif
     static RngEngine Reseed(RngEngine &eng)
@@ -407,15 +385,14 @@ namespace Grid {
       // MT implementation does not implement fast discard even though
       // in principle this is possible
       ////////////////////////////////////////////////
-      std::vector<int> gcoor;
-      int rank,o_idx,i_idx;
 
       // Everybody loops over global volume.
-      for(int gidx=0;gidx<_grid->_gsites;gidx++){
+      parallel_for(int gidx=0;gidx<_grid->_gsites;gidx++){
-
-	Skip(master_engine); // Skip to next RNG sequence
 
 	// Where is it?
+	int rank,o_idx,i_idx;
+	std::vector<int> gcoor;
+
 	_grid->GlobalIndexToGlobalCoor(gidx,gcoor);
 	_grid->GlobalCoorToRankIndex(rank,o_idx,i_idx,gcoor);
 
@@ -423,6 +400,7 @@ namespace Grid {
 	if( rank == _grid->ThisRank() ){
 	  int l_idx=generator_idx(o_idx,i_idx);
 	  _generators[l_idx] = master_engine;
+	  Skip(_generators[l_idx],gidx); // Skip to next RNG sequence
 	}
 
       }

lib/lattice/Lattice_transfer.h

@@ -50,26 +50,22 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
   ////////////////////////////////////////////////////////////////////////////////////////////
   template<class vobj> inline void pickCheckerboard(int cb,Lattice<vobj> &half,const Lattice<vobj> &full){
     half.checkerboard = cb;
-    int ssh=0;
+
-    //parallel_for
+    parallel_for(int ss=0;ss<full._grid->oSites();ss++){
-    for(int ss=0;ss<full._grid->oSites();ss++){
-      std::vector<int> coor;
       int cbos;
-      
+      std::vector<int> coor;
       full._grid->oCoorFromOindex(coor,ss);
       cbos=half._grid->CheckerBoard(coor);
       
       if (cbos==cb) {
+	int ssh=half._grid->oIndex(coor);
 	half._odata[ssh] = full._odata[ss];
-	ssh++;
       }
     }
   }
   template<class vobj> inline void setCheckerboard(Lattice<vobj> &full,const Lattice<vobj> &half){
     int cb = half.checkerboard;
-    int ssh=0;
+    parallel_for(int ss=0;ss<full._grid->oSites();ss++){
-    //parallel_for
-    for(int ss=0;ss<full._grid->oSites();ss++){
       std::vector<int> coor;
       int cbos;
 
@@ -77,8 +73,8 @@ inline void subdivides(GridBase *coarse,GridBase *fine)
       cbos=half._grid->CheckerBoard(coor);
       
       if (cbos==cb) {
+	int ssh=half._grid->oIndex(coor);
 	full._odata[ss]=half._odata[ssh];
-	ssh++;
       }
     }
   }
@@ -109,8 +105,8 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
 
   coarseData=zero;
 
-  // Loop with a cache friendly loop ordering
+  // Loop over coars parallel, and then loop over fine associated with coarse.
-  for(int sf=0;sf<fine->oSites();sf++){
+  parallel_for(int sf=0;sf<fine->oSites();sf++){
 
     int sc;
     std::vector<int> coor_c(_ndimension);
@@ -119,8 +115,9 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
     for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
     Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
 
+PARALLEL_CRITICAL
     for(int i=0;i<nbasis;i++) {
-      
+
       coarseData._odata[sc](i)=coarseData._odata[sc](i)
 	+ innerProduct(Basis[i]._odata[sf],fineData._odata[sf]);
 
@@ -139,6 +136,7 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
   GridBase * coarse= coarseA._grid;
 
   fineZ.checkerboard=fineX.checkerboard;
+  assert(fineX.checkerboard==fineY.checkerboard);
   subdivides(coarse,fine); // require they map
   conformable(fineX,fineY);
   conformable(fineX,fineZ);
@@ -180,9 +178,10 @@ template<class vobj,class CComplex>
   GridBase *coarse(CoarseInner._grid);
   GridBase *fine  (fineX._grid);
 
-  Lattice<dotp> fine_inner(fine);
+  Lattice<dotp> fine_inner(fine); fine_inner.checkerboard = fineX.checkerboard;
   Lattice<dotp> coarse_inner(coarse);
 
+  // Precision promotion?
   fine_inner = localInnerProduct(fineX,fineY);
   blockSum(coarse_inner,fine_inner);
   parallel_for(int ss=0;ss<coarse->oSites();ss++){
@@ -193,7 +192,7 @@ template<class vobj,class CComplex>
 inline void blockNormalise(Lattice<CComplex> &ip,Lattice<vobj> &fineX)
 {
   GridBase *coarse = ip._grid;
-  Lattice<vobj> zz(fineX._grid); zz=zero;
+  Lattice<vobj> zz(fineX._grid); zz=zero; zz.checkerboard=fineX.checkerboard;
   blockInnerProduct(ip,fineX,fineX);
   ip = pow(ip,-0.5);
   blockZAXPY(fineX,ip,fineX,zz);
@@ -216,19 +215,25 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
     block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
   }
 
+  // Turn this around to loop threaded over sc and interior loop 
+  // over sf would thread better
   coarseData=zero;
-  for(int sf=0;sf<fine->oSites();sf++){
+  parallel_region {
-    
+
     int sc;
     std::vector<int> coor_c(_ndimension);
     std::vector<int> coor_f(_ndimension);
 
-    Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
+    parallel_for_internal(int sf=0;sf<fine->oSites();sf++){
-    for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
+    
-    Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
+      Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
-
+      for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
-    coarseData._odata[sc]=coarseData._odata[sc]+fineData._odata[sf];
+      Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
+      
+PARALLEL_CRITICAL
+      coarseData._odata[sc]=coarseData._odata[sc]+fineData._odata[sf];
 
+    }
   }
   return;
 }
@@ -238,7 +243,7 @@ inline void blockPick(GridBase *coarse,const Lattice<vobj> &unpicked,Lattice<vob
 {
   GridBase * fine = unpicked._grid;
 
-  Lattice<vobj> zz(fine);
+  Lattice<vobj> zz(fine); zz.checkerboard = unpicked.checkerboard;
   Lattice<iScalar<vInteger> > fcoor(fine);
 
   zz = zero;
@@ -303,20 +308,21 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
   }
 
   // Loop with a cache friendly loop ordering
-  for(int sf=0;sf<fine->oSites();sf++){
+  parallel_region {
-
     int sc;
     std::vector<int> coor_c(_ndimension);
     std::vector<int> coor_f(_ndimension);
 
-    Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
+    parallel_for_internal(int sf=0;sf<fine->oSites();sf++){
-    for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
-    Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
-
-    for(int i=0;i<nbasis;i++) {
-      if(i==0) fineData._odata[sf]=coarseData._odata[sc](i) * Basis[i]._odata[sf];
-      else     fineData._odata[sf]=fineData._odata[sf]+coarseData._odata[sc](i)*Basis[i]._odata[sf];
 
+      Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
+      for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
+      Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
+      
+      for(int i=0;i<nbasis;i++) {
+	if(i==0) fineData._odata[sf]=coarseData._odata[sc](i) * Basis[i]._odata[sf];
+	else     fineData._odata[sf]=fineData._odata[sf]+coarseData._odata[sc](i)*Basis[i]._odata[sf];
+      }
     }
   }
   return;
@@ -684,6 +690,302 @@ void precisionChange(Lattice<VobjOut> &out, const Lattice<VobjIn> &in){
     merge(out._odata[out_oidx], ptrs, 0);
   }
 }
- 
+
+////////////////////////////////////////////////////////////////////////////////
+// Communicate between grids
+////////////////////////////////////////////////////////////////////////////////
+///////////////////////////////////////////////////////////////////////////////////////////////////////////
+// SIMPLE CASE:
+///////////////////////////////////////////////////////////////////////////////////////////////////////////
+//
+// Mesh of nodes (2x2) ; subdivide to  1x1 subdivisions
+//
+// Lex ord:   
+//          N0 va0 vb0 vc0 vd0       N1 va1 vb1 vc1 vd1  
+//          N2 va2 vb2 vc2 vd2       N3 va3 vb3 vc3 vd3 
+//
+// Ratio = full[dim] / split[dim]
+//
+// For each dimension do an all to all; get Nvec -> Nvec / ratio
+//                                          Ldim -> Ldim * ratio
+//                                          LocalVol -> LocalVol * ratio
+// full AllToAll(0)
+//          N0 va0 vb0 va1 vb1       N1 vc0 vd0 vc1 vd1   
+//          N2 va2 vb2 va3 vb3       N3 vc2 vd2 vc3 vd3 
+//
+// REARRANGE
+//          N0 va01 vb01      N1 vc01 vd01
+//          N2 va23 vb23      N3 vc23 vd23
+//
+// full AllToAll(1)           // Not what is wanted. FIXME
+//          N0 va01 va23      N1 vc01 vc23 
+//          N2 vb01 vb23      N3 vd01 vd23
+// 
+// REARRANGE
+//          N0 va0123      N1 vc0123
+//          N2 vb0123      N3 vd0123
+//
+// Must also rearrange data to get into the NEW lex order of grid at each stage. Some kind of "insert/extract".
+// NB: Easiest to programme if keep in lex order.
+/*
+ *  Let chunk = (fvol*nvec)/sP be size of a chunk.         ( Divide lexico vol * nvec into fP/sP = M chunks )
+ *  
+ *  2nd A2A (over sP nodes; subdivide the fP into sP chunks of M)
+ * 
+ *     node 0     1st chunk of node 0M..(1M-1); 2nd chunk of node 0M..(1M-1)..   data chunk x M x sP = fL / sP * M * sP = fL * M growth
+ *     node 1     1st chunk of node 1M..(2M-1); 2nd chunk of node 1M..(2M-1)..
+ *     node 2     1st chunk of node 2M..(3M-1); 2nd chunk of node 2M..(3M-1)..
+ *     node 3     1st chunk of node 3M..(3M-1); 2nd chunk of node 2M..(3M-1)..
+ *  etc...
+ */
+template<class Vobj>
+void Grid_split(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
+{
+  typedef typename Vobj::scalar_object Sobj;
+
+  int full_vecs   = full.size();
+
+  assert(full_vecs>=1);
+
+  GridBase * full_grid = full[0]._grid;
+  GridBase *split_grid = split._grid;
+
+  int       ndim  = full_grid->_ndimension;
+  int  full_nproc = full_grid->_Nprocessors;
+  int split_nproc =split_grid->_Nprocessors;
+
+  ////////////////////////////////
+  // Checkerboard management
+  ////////////////////////////////
+  int cb = full[0].checkerboard;
+  split.checkerboard = cb;
+
+  //////////////////////////////
+  // Checks
+  //////////////////////////////
+  assert(full_grid->_ndimension==split_grid->_ndimension);
+  for(int n=0;n<full_vecs;n++){
+    assert(full[n].checkerboard == cb);
+    for(int d=0;d<ndim;d++){
+      assert(full[n]._grid->_gdimensions[d]==split._grid->_gdimensions[d]);
+      assert(full[n]._grid->_fdimensions[d]==split._grid->_fdimensions[d]);
+    }
+  }
+
+  int   nvector   =full_nproc/split_nproc; 
+  assert(nvector*split_nproc==full_nproc);
+  assert(nvector == full_vecs);
+
+  std::vector<int> ratio(ndim);
+  for(int d=0;d<ndim;d++){
+    ratio[d] = full_grid->_processors[d]/ split_grid->_processors[d];
+  }
+
+  uint64_t lsites = full_grid->lSites();
+  uint64_t     sz = lsites * nvector;
+  std::vector<Sobj> tmpdata(sz);
+  std::vector<Sobj> alldata(sz);
+  std::vector<Sobj> scalardata(lsites); 
+
+  for(int v=0;v<nvector;v++){
+    unvectorizeToLexOrdArray(scalardata,full[v]);    
+    parallel_for(int site=0;site<lsites;site++){
+      alldata[v*lsites+site] = scalardata[site];
+    }
+  }
+
+  int nvec = nvector; // Counts down to 1 as we collapse dims
+  std::vector<int> ldims = full_grid->_ldimensions;
+
+  for(int d=ndim-1;d>=0;d--){
+
+    if ( ratio[d] != 1 ) {
+
+      full_grid ->AllToAll(d,alldata,tmpdata);
+      if ( split_grid->_processors[d] > 1 ) {
+	alldata=tmpdata;
+	split_grid->AllToAll(d,alldata,tmpdata);
+      }
+
+      auto rdims = ldims; 
+      auto     M = ratio[d];
+      auto rsites= lsites*M;// increases rsites by M
+      nvec      /= M;       // Reduce nvec by subdivision factor
+      rdims[d]  *= M;       // increase local dim by same factor
+
+      int sP =   split_grid->_processors[d];
+      int fP =    full_grid->_processors[d];
+
+      int fvol   = lsites;
+      
+      int chunk  = (nvec*fvol)/sP;          assert(chunk*sP == nvec*fvol);
+
+      // Loop over reordered data post A2A
+      parallel_for(int c=0;c<chunk;c++){
+	std::vector<int> coor(ndim);
+	for(int m=0;m<M;m++){
+	  for(int s=0;s<sP;s++){
+	    
+	    // addressing; use lexico
+	    int lex_r;
+	    uint64_t lex_c        = c+chunk*m+chunk*M*s;
+	    uint64_t lex_fvol_vec = c+chunk*s;
+	    uint64_t lex_fvol     = lex_fvol_vec%fvol;
+	    uint64_t lex_vec      = lex_fvol_vec/fvol;
+
+	    // which node sets an adder to the coordinate
+	    Lexicographic::CoorFromIndex(coor, lex_fvol, ldims);	  
+	    coor[d] += m*ldims[d];
+	    Lexicographic::IndexFromCoor(coor, lex_r, rdims);	  
+	    lex_r += lex_vec * rsites;
+
+	    // LexicoFind coordinate & vector number within split lattice
+	    alldata[lex_r] = tmpdata[lex_c];
+
+	  }
+	}
+      }
+      ldims[d]*= ratio[d];
+      lsites  *= ratio[d];
+
+    }
+  }
+  vectorizeFromLexOrdArray(alldata,split);    
+}
+
+template<class Vobj>
+void Grid_split(Lattice<Vobj> &full,Lattice<Vobj>   & split)
+{
+  int nvector = full._grid->_Nprocessors / split._grid->_Nprocessors;
+  std::vector<Lattice<Vobj> > full_v(nvector,full._grid);
+  for(int n=0;n<nvector;n++){
+    full_v[n] = full;
+  }
+  Grid_split(full_v,split);
+}
+
+template<class Vobj>
+void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj>   & split)
+{
+  typedef typename Vobj::scalar_object Sobj;
+
+  int full_vecs   = full.size();
+
+  assert(full_vecs>=1);
+
+  GridBase * full_grid = full[0]._grid;
+  GridBase *split_grid = split._grid;
+
+  int       ndim  = full_grid->_ndimension;
+  int  full_nproc = full_grid->_Nprocessors;
+  int split_nproc =split_grid->_Nprocessors;
+
+  ////////////////////////////////
+  // Checkerboard management
+  ////////////////////////////////
+  int cb = full[0].checkerboard;
+  split.checkerboard = cb;
+
+  //////////////////////////////
+  // Checks
+  //////////////////////////////
+  assert(full_grid->_ndimension==split_grid->_ndimension);
+  for(int n=0;n<full_vecs;n++){
+    assert(full[n].checkerboard == cb);
+    for(int d=0;d<ndim;d++){
+      assert(full[n]._grid->_gdimensions[d]==split._grid->_gdimensions[d]);
+      assert(full[n]._grid->_fdimensions[d]==split._grid->_fdimensions[d]);
+    }
+  }
+
+  int   nvector   =full_nproc/split_nproc; 
+  assert(nvector*split_nproc==full_nproc);
+  assert(nvector == full_vecs);
+
+  std::vector<int> ratio(ndim);
+  for(int d=0;d<ndim;d++){
+    ratio[d] = full_grid->_processors[d]/ split_grid->_processors[d];
+  }
+
+  uint64_t lsites = full_grid->lSites();
+  uint64_t     sz = lsites * nvector;
+  std::vector<Sobj> tmpdata(sz);
+  std::vector<Sobj> alldata(sz);
+  std::vector<Sobj> scalardata(lsites); 
+
+  unvectorizeToLexOrdArray(alldata,split);    
+
+  /////////////////////////////////////////////////////////////////
+  // Start from split grid and work towards full grid
+  /////////////////////////////////////////////////////////////////
+
+  int nvec = 1;
+  uint64_t rsites        = split_grid->lSites();
+  std::vector<int> rdims = split_grid->_ldimensions;
+
+  for(int d=0;d<ndim;d++){
+
+    if ( ratio[d] != 1 ) {
+
+      auto     M = ratio[d];
+
+      int sP =   split_grid->_processors[d];
+      int fP =    full_grid->_processors[d];
+      
+      auto ldims = rdims;  ldims[d]  /= M;  // Decrease local dims by same factor
+      auto lsites= rsites/M;                // Decreases rsites by M
+      
+      int fvol   = lsites;
+      int chunk  = (nvec*fvol)/sP;          assert(chunk*sP == nvec*fvol);
+	
+      {
+	// Loop over reordered data post A2A
+	parallel_for(int c=0;c<chunk;c++){
+	  std::vector<int> coor(ndim);
+	  for(int m=0;m<M;m++){
+	    for(int s=0;s<sP;s++){
+
+	      // addressing; use lexico
+	      int lex_r;
+	      uint64_t lex_c = c+chunk*m+chunk*M*s;
+	      uint64_t lex_fvol_vec = c+chunk*s;
+	      uint64_t lex_fvol     = lex_fvol_vec%fvol;
+	      uint64_t lex_vec      = lex_fvol_vec/fvol;
+	      
+	      // which node sets an adder to the coordinate
+	      Lexicographic::CoorFromIndex(coor, lex_fvol, ldims);	  
+	      coor[d] += m*ldims[d];
+	      Lexicographic::IndexFromCoor(coor, lex_r, rdims);	  
+	      lex_r += lex_vec * rsites;
+	      
+	      // LexicoFind coordinate & vector number within split lattice
+	      tmpdata[lex_c] = alldata[lex_r];
+	    }
+	  }
+	}
+      }
+
+      if ( split_grid->_processors[d] > 1 ) {
+	split_grid->AllToAll(d,tmpdata,alldata);
+	tmpdata=alldata;
+      }
+      full_grid ->AllToAll(d,tmpdata,alldata);
+      rdims[d]/= M;
+      rsites  /= M;
+      nvec    *= M;       // Increase nvec by subdivision factor
+    }
+  }
+
+  lsites = full_grid->lSites();
+  for(int v=0;v<nvector;v++){
+    //    assert(v<full.size());
+    parallel_for(int site=0;site<lsites;site++){
+      //      assert(v*lsites+site < alldata.size());
+      scalardata[site] = alldata[v*lsites+site];
+    }
+    vectorizeFromLexOrdArray(scalardata,full[v]);    
+  }
+}
+
 }
 #endif

lib/log/Log.cc

@@ -50,7 +50,7 @@ namespace Grid {
     return (status==0) ? res.get() : name ;
   }
   
-GridStopWatch Logger::StopWatch;
+GridStopWatch Logger::GlobalStopWatch;
 int Logger::timestamp;
 std::ostream Logger::devnull(0);
 
@@ -59,13 +59,15 @@ void GridLogTimestamp(int on){
 }
 
 Colours GridLogColours(0);
-GridLogger GridLogError(1, "Error", GridLogColours, "RED");
+GridLogger GridLogIRL    (1, "IRL"   , GridLogColours, "NORMAL");
+GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
+GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");
 GridLogger GridLogWarning(1, "Warning", GridLogColours, "YELLOW");
 GridLogger GridLogMessage(1, "Message", GridLogColours, "NORMAL");
-GridLogger GridLogDebug(1, "Debug", GridLogColours, "PURPLE");
+GridLogger GridLogDebug  (1, "Debug", GridLogColours, "PURPLE");
 GridLogger GridLogPerformance(1, "Performance", GridLogColours, "GREEN");
-GridLogger GridLogIterative(1, "Iterative", GridLogColours, "BLUE");
+GridLogger GridLogIterative  (1, "Iterative", GridLogColours, "BLUE");
-GridLogger GridLogIntegrator(1, "Integrator", GridLogColours, "BLUE");
+GridLogger GridLogIntegrator (1, "Integrator", GridLogColours, "BLUE");
 
 void GridLogConfigure(std::vector<std::string> &logstreams) {
   GridLogError.Active(0);

lib/log/Log.h

@@ -85,12 +85,16 @@ class Logger {
 protected:
   Colours &Painter;
   int active;
+  int timing_mode;
+  int topWidth{-1};
   static int timestamp;
   std::string name, topName;
   std::string COLOUR;
 
 public:
-  static GridStopWatch StopWatch;
+  static GridStopWatch GlobalStopWatch;
+  GridStopWatch         LocalStopWatch;
+  GridStopWatch *StopWatch;
   static std::ostream devnull;
 
   std::string background() {return Painter.colour["NORMAL"];}
@@ -101,22 +105,44 @@ public:
     name(nm),
     topName(topNm),
     Painter(col_class),
-    COLOUR(col) {} ;
+    timing_mode(0),
+    COLOUR(col) 
+    {
+      StopWatch = & GlobalStopWatch;
+    };
   
   void Active(int on) {active = on;};
   int  isActive(void) {return active;};
   static void Timestamp(int on) {timestamp = on;};
-  
+  void Reset(void) { 
+    StopWatch->Reset(); 
+    StopWatch->Start(); 
+  }
+  void TimingMode(int on) { 
+    timing_mode = on; 
+    if(on) { 
+      StopWatch = &LocalStopWatch;
+      Reset(); 
+    }
+  }
+  void setTopWidth(const int w) {topWidth = w;}
+
   friend std::ostream& operator<< (std::ostream& stream, Logger& log){
 
     if ( log.active ) {
-      stream << log.background()<< std::setw(8) << std::left << log.topName << log.background()<< " : ";
+      stream << log.background()<<  std::left;
-      stream << log.colour() << std::setw(10) << std::left << log.name << log.background() << " : ";
+      if (log.topWidth > 0)
+      {
+        stream << std::setw(log.topWidth);
+      }
+      stream << log.topName << log.background()<< " : ";
+      stream << log.colour() <<  std::left << log.name << log.background() << " : ";
       if ( log.timestamp ) {
-	StopWatch.Stop();
+	log.StopWatch->Stop();
-	GridTime now = StopWatch.Elapsed();
+	GridTime now = log.StopWatch->Elapsed();
-	StopWatch.Start();
+	if ( log.timing_mode==1 ) log.StopWatch->Reset();
-	stream << log.evidence()<< now << log.background() << " : " ;
+	log.StopWatch->Start();
+	stream << log.evidence()<< std::setw(6)<<now << log.background() << " : " ;
       }
       stream << log.colour();
       return stream;
@@ -135,6 +161,8 @@ public:
 
 void GridLogConfigure(std::vector<std::string> &logstreams);
 
+extern GridLogger GridLogIRL;
+extern GridLogger GridLogSolver;
 extern GridLogger GridLogError;
 extern GridLogger GridLogWarning;
 extern GridLogger GridLogMessage;

lib/parallelIO/BinaryIO.h

@@ -261,7 +261,7 @@ class BinaryIO {
 			      GridBase *grid,
 			      std::vector<fobj> &iodata,
 			      std::string file,
-			      int offset,
+			      Integer offset,
 			      const std::string &format, int control,
 			      uint32_t &nersc_csum,
 			      uint32_t &scidac_csuma,
@@ -356,7 +356,7 @@ class BinaryIO {
 
       if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
 #ifdef USE_MPI_IO
-	std::cout<< GridLogMessage<< "MPI read I/O "<< file<< std::endl;
+	std::cout<< GridLogMessage<<"IOobject: MPI read I/O "<< file<< std::endl;
 	ierr=MPI_File_open(grid->communicator,(char *) file.c_str(), MPI_MODE_RDONLY, MPI_INFO_NULL, &fh);    assert(ierr==0);
 	ierr=MPI_File_set_view(fh, disp, mpiObject, fileArray, "native", MPI_INFO_NULL);    assert(ierr==0);
 	ierr=MPI_File_read_all(fh, &iodata[0], 1, localArray, &status);    assert(ierr==0);
@@ -367,7 +367,7 @@ class BinaryIO {
 	assert(0);
 #endif
       } else {
-        std::cout << GridLogMessage << "C++ read I/O " << file << " : "
+	std::cout << GridLogMessage <<"IOobject: C++ read I/O " << file << " : "
                   << iodata.size() * sizeof(fobj) << " bytes" << std::endl;
         std::ifstream fin;
         fin.open(file, std::ios::binary | std::ios::in);
@@ -413,9 +413,9 @@ class BinaryIO {
       timer.Start();
       if ( (control & BINARYIO_LEXICOGRAPHIC) && (nrank > 1) ) {
 #ifdef USE_MPI_IO
-        std::cout << GridLogMessage << "MPI write I/O " << file << std::endl;
+        std::cout << GridLogMessage <<"IOobject: MPI write I/O " << file << std::endl;
         ierr = MPI_File_open(grid->communicator, (char *)file.c_str(), MPI_MODE_RDWR | MPI_MODE_CREATE, MPI_INFO_NULL, &fh);
-        std::cout << GridLogMessage << "Checking for errors" << std::endl;
+	//        std::cout << GridLogMessage << "Checking for errors" << std::endl;
         if (ierr != MPI_SUCCESS)
         {
           char error_string[BUFSIZ];
@@ -444,48 +444,56 @@ class BinaryIO {
 	assert(0);
 #endif
       } else { 
+
+        std::cout << GridLogMessage << "IOobject: C++ write I/O " << file << " : "
+                  << iodata.size() * sizeof(fobj) << " bytes" << std::endl;
         
 	std::ofstream fout; 
-  fout.exceptions ( std::fstream::failbit | std::fstream::badbit );
+	fout.exceptions ( std::fstream::failbit | std::fstream::badbit );
-  try {
+	try {
-    fout.open(file,std::ios::binary|std::ios::out|std::ios::in);
+	  fout.open(file,std::ios::binary|std::ios::out|std::ios::in);
-  } catch (const std::fstream::failure& exc) {
+	} catch (const std::fstream::failure& exc) {
-    std::cout << GridLogError << "Error in opening the file " << file << " for output" <<std::endl;
+	  std::cout << GridLogError << "Error in opening the file " << file << " for output" <<std::endl;
-    std::cout << GridLogError << "Exception description: " << exc.what() << std::endl;
+	  std::cout << GridLogError << "Exception description: " << exc.what() << std::endl;
-    std::cout << GridLogError << "Probable cause: wrong path, inaccessible location "<< std::endl;
+	  std::cout << GridLogError << "Probable cause: wrong path, inaccessible location "<< std::endl;
-    #ifdef USE_MPI_IO
+#ifdef USE_MPI_IO
-    MPI_Abort(MPI_COMM_WORLD,1);
+	  MPI_Abort(MPI_COMM_WORLD,1);
-    #else
+#else
-    exit(1);
+	  exit(1);
-    #endif
+#endif
-  }
+	}
-	std::cout << GridLogMessage<< "C++ write I/O "<< file<<" : "
+	
-		        << iodata.size()*sizeof(fobj)<<" bytes"<<std::endl;
+	if ( control & BINARYIO_MASTER_APPEND )  {
-	
+	  try {
-  if ( control & BINARYIO_MASTER_APPEND )  {
+	    fout.seekp(0,fout.end);
-	  fout.seekp(0,fout.end);
+	  } catch (const std::fstream::failure& exc) {
-	} else {
+	    std::cout << "Exception in seeking file end " << file << std::endl;
-	  fout.seekp(offset+myrank*lsites*sizeof(fobj));
+	  }
+	} else {
+	  try { 
+	    fout.seekp(offset+myrank*lsites*sizeof(fobj));
+	  } catch (const std::fstream::failure& exc) {
+	    std::cout << "Exception in seeking file " << file <<" offset "<< offset << std::endl;
+	  }
 	}
-  
-  try {
-  	fout.write((char *)&iodata[0],iodata.size()*sizeof(fobj));//assert( fout.fail()==0);
-  }
-  catch (const std::fstream::failure& exc) {
-    std::cout << "Exception in writing file " << file << std::endl;
-    std::cout << GridLogError << "Exception description: "<< exc.what() << std::endl;
-    #ifdef USE_MPI_IO
-    MPI_Abort(MPI_COMM_WORLD,1);
-    #else
-    exit(1);
-    #endif
-  }
 
+	try {
+	  fout.write((char *)&iodata[0],iodata.size()*sizeof(fobj));//assert( fout.fail()==0);
+	}
+	catch (const std::fstream::failure& exc) {
+	  std::cout << "Exception in writing file " << file << std::endl;
+	  std::cout << GridLogError << "Exception description: "<< exc.what() << std::endl;
+#ifdef USE_MPI_IO
+	  MPI_Abort(MPI_COMM_WORLD,1);
+#else
+	  exit(1);
+#endif
+	}
 	fout.close();
-  }
+      }
-  timer.Stop();
+      timer.Stop();
-  }
+    }
-
+    
     std::cout<<GridLogMessage<<"IOobject: ";
     if ( control & BINARYIO_READ) std::cout << " read  ";
     else                          std::cout << " write ";
@@ -515,7 +523,7 @@ class BinaryIO {
   static inline void readLatticeObject(Lattice<vobj> &Umu,
 				       std::string file,
 				       munger munge,
-				       int offset,
+				       Integer offset,
 				       const std::string &format,
 				       uint32_t &nersc_csum,
 				       uint32_t &scidac_csuma,
@@ -552,7 +560,7 @@ class BinaryIO {
     static inline void writeLatticeObject(Lattice<vobj> &Umu,
 					  std::string file,
 					  munger munge,
-					  int offset,
+					  Integer offset,
 					  const std::string &format,
 					  uint32_t &nersc_csum,
 					  uint32_t &scidac_csuma,
@@ -589,7 +597,7 @@ class BinaryIO {
   static inline void readRNG(GridSerialRNG &serial,
 			     GridParallelRNG &parallel,
 			     std::string file,
-			     int offset,
+			     Integer offset,
 			     uint32_t &nersc_csum,
 			     uint32_t &scidac_csuma,
 			     uint32_t &scidac_csumb)
@@ -651,7 +659,7 @@ class BinaryIO {
   static inline void writeRNG(GridSerialRNG &serial,
 			      GridParallelRNG &parallel,
 			      std::string file,
-			      int offset,
+			      Integer offset,
 			      uint32_t &nersc_csum,
 			      uint32_t &scidac_csuma,
 			      uint32_t &scidac_csumb)