snn_manager.cpp

Go to the documentation of this file.

/* * Copyright (c) 2016 Regents of the University of California. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
*
* 1. Redistributions of source code must retain the above copyright
*    notice, this list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright
*    notice, this list of conditions and the following disclaimer in the
*    documentation and/or other materials provided with the distribution.
*
* 3. The names of its contributors may not be used to endorse or promote
*    products derived from this software without specific prior written
*    permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
* CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
* NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
* *********************************************************************************************** *
* CARLsim
* created by: (MDR) Micah Richert, (JN) Jayram M. Nageswaran
* maintained by:
* (MA) Mike Avery <averym@uci.edu>
* (MB) Michael Beyeler <mbeyeler@uci.edu>,
* (KDC) Kristofor Carlson <kdcarlso@uci.edu>
* (TSC) Ting-Shuo Chou <tingshuc@uci.edu>
* (HK) Hirak J Kashyap <kashyaph@uci.edu>
*
* CARLsim v1.0: JM, MDR
* CARLsim v2.0/v2.1/v2.2: JM, MDR, MA, MB, KDC
* CARLsim3: MB, KDC, TSC
* CARLsim4: TSC, HK
* CARLsim5: HK, JX, KC
*
* CARLsim available from http://socsci.uci.edu/~jkrichma/CARLsim/
* Ver 12/31/2016
*/
 
#include <snn.h>
#include <sstream>
#include <algorithm>
 
#include <connection_monitor.h>
#include <connection_monitor_core.h>
#include <spike_monitor.h>
#include <spike_monitor_core.h>
#include <group_monitor.h>
#include <group_monitor_core.h>
#include <neuron_monitor.h>
#include <neuron_monitor_core.h>
 
#include <spike_buffer.h>
#include <error_code.h>
 
// \FIXME what are the following for? why were they all the way at the bottom of this file?
 
#define COMPACTION_ALIGNMENT_PRE  16
#define COMPACTION_ALIGNMENT_POST 0
 
 
 
// TODO: consider moving unsafe computations out of constructor
SNN::SNN(const std::string& name, SimMode preferredSimMode, LoggerMode loggerMode, int randSeed)
                    : networkName_(name), preferredSimMode_(preferredSimMode), loggerMode_(loggerMode),
                      randSeed_(SNN::setRandSeed(randSeed)) // all of these are const
{
    // move all unsafe operations out of constructor
    SNNinit();
}
 
// destructor
SNN::~SNN() {
    if (!simulatorDeleted)
        deleteObjects();
}
 
 
// make from each neuron in grpId1 to 'numPostSynapses' neurons in grpId2
short int SNN::connect(int grpId1, int grpId2, const std::string& _type, float initWt, float maxWt, float prob,
                        uint8_t minDelay, uint8_t maxDelay, RadiusRF radius,
                        float _mulSynFast, float _mulSynSlow, bool synWtType) {
                        //const std::string& wtType
    int retId=-1;
    assert(grpId1 < numGroups);
    assert(grpId2 < numGroups);
    assert(minDelay <= maxDelay);
    assert(!isPoissonGroup(grpId2));
 
    //* \deprecated Do these ramp thingies still work?
//    bool useRandWts = (wtType.find("random") != std::string::npos);
//    bool useRampDownWts = (wtType.find("ramp-down") != std::string::npos);
//    bool useRampUpWts = (wtType.find("ramp-up") != std::string::npos);
//    uint32_t connProp = SET_INITWTS_RANDOM(useRandWts)
//      | SET_CONN_PRESENT(1)
//      | SET_FIXED_PLASTIC(synWtType)
//      | SET_INITWTS_RAMPUP(useRampUpWts)
//      | SET_INITWTS_RAMPDOWN(useRampDownWts);
    uint32_t connProp = SET_CONN_PRESENT(1) | SET_FIXED_PLASTIC(synWtType);
 
    Grid3D szPre = getGroupGrid3D(grpId1);
    Grid3D szPost = getGroupGrid3D(grpId2);
 
    // initialize configuration of a connection
    ConnectConfig connConfig;
 
    connConfig.grpSrc   = grpId1;
    connConfig.grpDest  = grpId2;
    connConfig.initWt   = initWt;
    connConfig.maxWt    = maxWt;
    connConfig.maxDelay = maxDelay;
    connConfig.minDelay = minDelay;
//      newInfo->radX             = (radX<0) ? MAX(szPre.x,szPost.x) : radX; // <0 means full connectivity, so the
//      newInfo->radY             = (radY<0) ? MAX(szPre.y,szPost.y) : radY; // effective group size is Grid3D.x. Grab
//      newInfo->radZ             = (radZ<0) ? MAX(szPre.z,szPost.z) : radZ; // the larger of pre / post to connect all
    connConfig.connRadius = radius;
    connConfig.mulSynFast      = _mulSynFast;
    connConfig.mulSynSlow      = _mulSynSlow;
    connConfig.connProp        = connProp;
    connConfig.connProbability = prob;
    connConfig.type            = CONN_UNKNOWN;
    connConfig.connectionMonitorId = -1;
    connConfig.connId = -1;
    connConfig.conn = NULL;
    connConfig.numberOfConnections = 0;
 
    if ( _type.find("random") != std::string::npos) {
        connConfig.type = CONN_RANDOM;
    }
    //so you're setting the size to be prob*Number of synapses in group info + some standard deviation ...
    else if ( _type.find("full-no-direct") != std::string::npos) {
        connConfig.type = CONN_FULL_NO_DIRECT;
    }
    else if ( _type.find("full") != std::string::npos) {
        connConfig.type = CONN_FULL;
    }
    else if ( _type.find("one-to-one") != std::string::npos) {
        connConfig.type = CONN_ONE_TO_ONE;
    } else if ( _type.find("gaussian") != std::string::npos) {
        connConfig.type   = CONN_GAUSSIAN;
    } else {
        KERNEL_ERROR("Invalid connection type (should be 'random', 'full', 'one-to-one', 'full-no-direct', or 'gaussian')");
        exitSimulation(-1);
    }
 
    // assign connection id
    assert(connConfig.connId == -1);
    connConfig.connId = numConnections;
 
    KERNEL_DEBUG("CONNECT SETUP: connId=%d, mulFast=%f, mulSlow=%f", connConfig.connId, connConfig.mulSynFast, connConfig.mulSynSlow);
 
    // store the configuration of a connection
    connectConfigMap[numConnections] = connConfig; // connConfig.connId == numConnections
 
    assert(numConnections < MAX_CONN_PER_SNN);  // make sure we don't overflow connId
    numConnections++;
 
    return (numConnections - 1);
}
 
// make custom connections from grpId1 to grpId2
short int SNN::connect(int grpId1, int grpId2, ConnectionGeneratorCore* conn, float _mulSynFast, float _mulSynSlow,
                        bool synWtType) {
    int retId=-1;
 
    assert(grpId1 < numGroups);
    assert(grpId2 < numGroups);
 
    // initialize the configuration of a connection
    ConnectConfig connConfig;
 
    connConfig.grpSrc   = grpId1;
    connConfig.grpDest  = grpId2;
    connConfig.initWt     = 0.0f;
    connConfig.maxWt      = 0.0f;
    connConfig.maxDelay = MAX_SYN_DELAY;
    connConfig.minDelay = 1;
    connConfig.mulSynFast = _mulSynFast;
    connConfig.mulSynSlow = _mulSynSlow;
    connConfig.connProp = SET_CONN_PRESENT(1) | SET_FIXED_PLASTIC(synWtType);
    connConfig.type = CONN_USER_DEFINED;
    connConfig.conn = conn;
    connConfig.connectionMonitorId = -1;
    connConfig.connId = -1;
    connConfig.numberOfConnections = 0;
 
    // assign a connection id
    assert(connConfig.connId == -1);
    connConfig.connId = numConnections;
 
    // store the configuration of a connection
    connectConfigMap[numConnections] = connConfig; // connConfig.connId == numConnections
 
    assert(numConnections < MAX_CONN_PER_SNN);  // make sure we don't overflow connId
    numConnections++;
 
    return (numConnections - 1);
}
 
// make a compartmental connection between two groups
short int SNN::connectCompartments(int grpIdLower, int grpIdUpper) {
    assert(grpIdLower >= 0 && grpIdLower < numGroups);
    assert(grpIdUpper >= 0 && grpIdUpper < numGroups);
    assert(grpIdLower != grpIdUpper);
    assert(!isPoissonGroup(grpIdLower));
    assert(!isPoissonGroup(grpIdUpper));
 
    // the two groups must be located on the same partition
    assert(groupConfigMap[grpIdLower].preferredNetId == groupConfigMap[grpIdUpper].preferredNetId);
 
    // this flag must be set if any compartmental connections exist
    // note that grpId.withCompartments is not necessarily set just yet, this will be done in
    // CpuSNN::setCompartmentParameters
    sim_with_compartments = true;
 
    compConnectConfig compConnConfig;
 
    compConnConfig.grpSrc = grpIdLower;
    compConnConfig.grpDest = grpIdUpper;
    compConnConfig.connId = -1;
 
    // assign a connection id
    assert(compConnConfig.connId == -1);
    compConnConfig.connId = numCompartmentConnections;
 
    // store the configuration of a connection
    compConnectConfigMap[numCompartmentConnections] = compConnConfig;
 
    numCompartmentConnections++;
 
    return (numCompartmentConnections - 1);
}
 
// create group of Izhikevich neurons
// use int for nNeur to avoid arithmetic underflow
int SNN::createGroup(const std::string& grpName, const Grid3D& grid, int neurType, int preferredPartition, ComputingBackend preferredBackend) {
    assert(grid.numX * grid.numY * grid.numZ > 0);
    assert(neurType >= 0);
    assert(numGroups < MAX_GRP_PER_SNN);
 
    if ( (!(neurType & TARGET_AMPA) && !(neurType & TARGET_NMDA) &&
          !(neurType & TARGET_GABAa) && !(neurType & TARGET_GABAb)) || (neurType & POISSON_NEURON)) {
        KERNEL_ERROR("Invalid type using createGroup... Cannot create poisson generators here.");
        exitSimulation(1);
    }
 
    // initialize group configuration
    GroupConfig grpConfig;
    GroupConfigMD grpConfigMD;
 
    //All groups are non-compartmental by default
    grpConfig.withCompartments = false;
 
    // init parameters of neural group size and location
    grpConfig.grpName = grpName;
    grpConfig.type = neurType;
    grpConfig.numN = grid.N;
 
    grpConfig.isSpikeGenerator = false;
    grpConfig.grid = grid;
    grpConfig.isLIF = false;
 
    if (preferredPartition == ANY) {
        grpConfig.preferredNetId = ANY;
    } else if (preferredBackend == CPU_CORES) {
        grpConfig.preferredNetId = preferredPartition + CPU_RUNTIME_BASE;
    } else {
        grpConfig.preferredNetId = preferredPartition + GPU_RUNTIME_BASE;
    }
 
    // assign a global group id
    grpConfigMD.gGrpId = numGroups;
 
    // store the configuration of a group
    groupConfigMap[numGroups] = grpConfig; // numGroups == grpId
    groupConfigMDMap[numGroups] = grpConfigMD;
 
    assert(numGroups < MAX_GRP_PER_SNN); // make sure we don't overflow connId
    numGroups++;
 
    return grpConfigMD.gGrpId;
}
 
// create group of LIF neurons
// use int for nNeur to avoid arithmetic underflow
int SNN::createGroupLIF(const std::string& grpName, const Grid3D& grid, int neurType, int preferredPartition, ComputingBackend preferredBackend) {
    assert(grid.numX * grid.numY * grid.numZ > 0);
    assert(neurType >= 0);
    assert(numGroups < MAX_GRP_PER_SNN);
 
    if ( (!(neurType & TARGET_AMPA) && !(neurType & TARGET_NMDA) &&
          !(neurType & TARGET_GABAa) && !(neurType & TARGET_GABAb)) || (neurType & POISSON_NEURON)) {
        KERNEL_ERROR("Invalid type using createGroup... Cannot create poisson generators here.");
        exitSimulation(1);
    }
 
    // initialize group configuration
    GroupConfig grpConfig;
    GroupConfigMD grpConfigMD;
 
    // init parameters of neural group size and location
    grpConfig.grpName = grpName;
    grpConfig.type = neurType;
    grpConfig.numN = grid.N;
 
    grpConfig.isLIF = true;
    grpConfig.isSpikeGenerator = false;
    grpConfig.grid = grid;
 
    if (preferredPartition == ANY) {
        grpConfig.preferredNetId = ANY;
    } else if (preferredBackend == CPU_CORES) {
        grpConfig.preferredNetId = preferredPartition + CPU_RUNTIME_BASE;
    } else {
        grpConfig.preferredNetId = preferredPartition + GPU_RUNTIME_BASE;
    }
 
    // assign a global group id
    grpConfigMD.gGrpId = numGroups;
 
    // store the configuration of a group
    groupConfigMap[numGroups] = grpConfig; // numGroups == grpId
    groupConfigMDMap[numGroups] = grpConfigMD;
 
    assert(numGroups < MAX_GRP_PER_SNN); // make sure we don't overflow connId
    numGroups++;
 
    return grpConfigMD.gGrpId;
}
 
// create spike generator group
// use int for nNeur to avoid arithmetic underflow
int SNN::createSpikeGeneratorGroup(const std::string& grpName, const Grid3D& grid, int neurType, int preferredPartition, ComputingBackend preferredBackend) {
    assert(grid.numX * grid.numY * grid.numZ > 0);
    assert(neurType >= 0);
    assert(numGroups < MAX_GRP_PER_SNN);
 
    // initialize group configuration
    GroupConfig grpConfig;
    GroupConfigMD grpConfigMD;
 
    //All groups are non-compartmental by default  FIXME:IS THIS NECESSARY?
    grpConfig.withCompartments = false;
 
    // init parameters of neural group size and location
    grpConfig.grpName = grpName;
    grpConfig.type = neurType | POISSON_NEURON;
    grpConfig.numN = grid.N;
    grpConfig.isSpikeGenerator = true;
    grpConfig.grid = grid;
    grpConfig.isLIF = false;
 
    if (preferredPartition == ANY) {
        grpConfig.preferredNetId = ANY;
    }
    else if (preferredBackend == CPU_CORES) {
        grpConfig.preferredNetId = preferredPartition + CPU_RUNTIME_BASE;
    }
    else {
        grpConfig.preferredNetId = preferredPartition + GPU_RUNTIME_BASE;
    }
 
    // assign a global group id
    grpConfigMD.gGrpId = numGroups;
 
    // store the configuration of a group
    groupConfigMap[numGroups] = grpConfig;
    groupConfigMDMap[numGroups] = grpConfigMD;
 
    assert(numGroups < MAX_GRP_PER_SNN); // make sure we don't overflow connId
    numGroups++;
    numSpikeGenGrps++;
 
    return grpConfigMD.gGrpId;
}
 
void SNN::setCompartmentParameters(int gGrpId, float couplingUp, float couplingDown) {
    if (gGrpId == ALL) {
        for (int grpId = 0; grpId<numGroups; grpId++) {
            setCompartmentParameters(grpId, couplingUp, couplingDown);
        }
    }
    else {
        groupConfigMap[gGrpId].withCompartments = true;
        groupConfigMap[gGrpId].compCouplingUp = couplingUp;
        groupConfigMap[gGrpId].compCouplingDown = couplingDown;
        glbNetworkConfig.numComp += groupConfigMap[gGrpId].numN;
    }
}
 
 
// set conductance values for a simulation (custom values or disable conductances alltogether)
void SNN::setConductances(bool isSet, int tdAMPA, int trNMDA, int tdNMDA, int tdGABAa, int trGABAb, int tdGABAb) {
    if (isSet) {
        assert(tdAMPA>0); assert(tdNMDA>0); assert(tdGABAa>0); assert(tdGABAb>0);
        assert(trNMDA>=0); assert(trGABAb>=0); // 0 to disable rise times
        assert(trNMDA!=tdNMDA); assert(trGABAb!=tdGABAb); // singularity
    }
 
    // set conductances globally for all connections
    sim_with_conductances  |= isSet;
    dAMPA  = 1.0-1.0/tdAMPA;
    dNMDA  = 1.0-1.0/tdNMDA;
    dGABAa = 1.0-1.0/tdGABAa;
    dGABAb = 1.0-1.0/tdGABAb;
 
    if (trNMDA>0) {
        // use rise time for NMDA
        sim_with_NMDA_rise = true;
        rNMDA = 1.0-1.0/trNMDA;
 
        // compute max conductance under this model to scale it back to 1
        // otherwise the peak conductance will not be equal to the weight
        double tmax = (-tdNMDA*trNMDA*log(1.0*trNMDA/tdNMDA))/(tdNMDA-trNMDA); // t at which cond will be max
        sNMDA = 1.0/(exp(-tmax/tdNMDA)-exp(-tmax/trNMDA)); // scaling factor, 1 over max amplitude
        assert(!isinf(tmax) && !isnan(tmax) && tmax>=0);
        assert(!isinf(sNMDA) && !isnan(sNMDA) && sNMDA>0);
    }
 
    if (trGABAb>0) {
        // use rise time for GABAb
        sim_with_GABAb_rise = true;
        rGABAb = 1.0-1.0/trGABAb;
 
        // compute max conductance under this model to scale it back to 1
        // otherwise the peak conductance will not be equal to the weight
        double tmax = (-tdGABAb*trGABAb*log(1.0*trGABAb/tdGABAb))/(tdGABAb-trGABAb); // t at which cond will be max
        sGABAb = 1.0/(exp(-tmax/tdGABAb)-exp(-tmax/trGABAb)); // scaling factor, 1 over max amplitude
        assert(!isinf(tmax) && !isnan(tmax)); assert(!isinf(sGABAb) && !isnan(sGABAb) && sGABAb>0);
    }
 
    if (sim_with_conductances) {
        KERNEL_INFO("Running COBA mode:");
        KERNEL_INFO("  - AMPA decay time            = %5d ms", tdAMPA);
        KERNEL_INFO("  - NMDA rise time %s  = %5d ms", sim_with_NMDA_rise?"          ":"(disabled)", trNMDA);
        KERNEL_INFO("  - GABAa decay time           = %5d ms", tdGABAa);
        KERNEL_INFO("  - GABAb rise time %s = %5d ms", sim_with_GABAb_rise?"          ":"(disabled)",trGABAb);
        KERNEL_INFO("  - GABAb decay time           = %5d ms", tdGABAb);
    } else {
        KERNEL_INFO("Running CUBA mode (all synaptic conductances disabled)");
    }
}
 
// set homeostasis for group
void SNN::setHomeostasis(int gGrpId, bool isSet, float homeoScale, float avgTimeScale) {
    if (gGrpId == ALL) { // shortcut for all groups
        for(int grpId = 0; grpId < numGroups; grpId++) {
            setHomeostasis(grpId, isSet, homeoScale, avgTimeScale);
        }
    } else {
        // set conductances for a given group
        sim_with_homeostasis                      |= isSet;
        groupConfigMap[gGrpId].homeoConfig.WithHomeostasis    = isSet;
        groupConfigMap[gGrpId].homeoConfig.homeostasisScale   = homeoScale;
        groupConfigMap[gGrpId].homeoConfig.avgTimeScale       = avgTimeScale;
        groupConfigMap[gGrpId].homeoConfig.avgTimeScaleInv    = 1.0f / avgTimeScale;
        groupConfigMap[gGrpId].homeoConfig.avgTimeScaleDecay = (avgTimeScale * 1000.0f - 1.0f) / (avgTimeScale * 1000.0f);
 
        KERNEL_INFO("Homeostasis parameters %s for %d (%s):\thomeoScale: %f, avgTimeScale: %f",
                    isSet?"enabled":"disabled", gGrpId, groupConfigMap[gGrpId].grpName.c_str(), homeoScale, avgTimeScale);
    }
}
 
// set a homeostatic target firing rate (enforced through homeostatic synaptic scaling)
void SNN::setHomeoBaseFiringRate(int gGrpId, float baseFiring, float baseFiringSD) {
    if (gGrpId == ALL) { // shortcut for all groups
        for(int grpId = 0; grpId < numGroups; grpId++) {
            setHomeoBaseFiringRate(grpId, baseFiring, baseFiringSD);
        }
    } else {
        // set homeostatsis for a given group
        groupConfigMap[gGrpId].homeoConfig.baseFiring = baseFiring;
        groupConfigMap[gGrpId].homeoConfig.baseFiringSD = baseFiringSD;
 
        KERNEL_INFO("Homeostatic base firing rate set for %d (%s):\tbaseFiring: %3.3f, baseFiringStd: %3.3f",
                            gGrpId, groupConfigMap[gGrpId].grpName.c_str(), baseFiring, baseFiringSD);
    }
}
 
 
void SNN::setIntegrationMethod(integrationMethod_t method, int numStepsPerMs) {
    assert(numStepsPerMs >= 1 && numStepsPerMs <= 100);
    glbNetworkConfig.simIntegrationMethod = method;
    glbNetworkConfig.simNumStepsPerMs = numStepsPerMs;
    glbNetworkConfig.timeStep = 1.0f / numStepsPerMs;
}
 
// set Izhikevich parameters for group
void SNN::setNeuronParameters(int gGrpId, float izh_a, float izh_a_sd, float izh_b, float izh_b_sd,
                                float izh_c, float izh_c_sd, float izh_d, float izh_d_sd)
{
    assert(gGrpId >= -1);
    assert(izh_a_sd >= 0); assert(izh_b_sd >= 0); assert(izh_c_sd >= 0); assert(izh_d_sd >= 0);
 
    if (gGrpId == ALL) { // shortcut for all groups
        for(int grpId = 0; grpId < numGroups; grpId++) {
            setNeuronParameters(grpId, izh_a, izh_a_sd, izh_b, izh_b_sd, izh_c, izh_c_sd, izh_d, izh_d_sd);
        }
    } else {
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_a = izh_a;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_a_sd = izh_a_sd;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_b = izh_b;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_b_sd = izh_b_sd;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_c = izh_c;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_c_sd = izh_c_sd;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_d = izh_d;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_d_sd = izh_d_sd;
        groupConfigMap[gGrpId].withParamModel_9 = 0;
        groupConfigMap[gGrpId].isLIF = 0;
    }
}
 
// set (9) Izhikevich parameters for group
void SNN::setNeuronParameters(int gGrpId, float izh_C, float izh_C_sd, float izh_k, float izh_k_sd,
    float izh_vr, float izh_vr_sd, float izh_vt, float izh_vt_sd,
    float izh_a, float izh_a_sd, float izh_b, float izh_b_sd,
    float izh_vpeak, float izh_vpeak_sd, float izh_c, float izh_c_sd,
    float izh_d, float izh_d_sd)
{
    assert(gGrpId >= -1);
    assert(izh_C_sd >= 0); assert(izh_k_sd >= 0); assert(izh_vr_sd >= 0);
    assert(izh_vt_sd >= 0); assert(izh_a_sd >= 0); assert(izh_b_sd >= 0); assert(izh_vpeak_sd >= 0);
    assert(izh_c_sd >= 0); assert(izh_d_sd >= 0);
 
    if (gGrpId == ALL) { // shortcut for all groups
        for (int grpId = 0; grpId<numGroups; grpId++) {
            setNeuronParameters(grpId, izh_C, izh_C_sd, izh_k, izh_k_sd, izh_vr, izh_vr_sd, izh_vt, izh_vt_sd,
                izh_a, izh_a_sd, izh_b, izh_b_sd, izh_vpeak, izh_vpeak_sd, izh_c, izh_c_sd,
                izh_d, izh_d_sd);
        }
    }
    else {
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_a = izh_a;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_a_sd = izh_a_sd;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_b = izh_b;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_b_sd = izh_b_sd;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_c = izh_c;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_c_sd = izh_c_sd;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_d = izh_d;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_d_sd = izh_d_sd;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_C = izh_C;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_C_sd = izh_C_sd;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_k = izh_k;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_k_sd = izh_k_sd;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vr = izh_vr;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vr_sd = izh_vr_sd;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vt = izh_vt;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vt_sd = izh_vt_sd;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vpeak = izh_vpeak;
        groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vpeak_sd = izh_vpeak_sd;
        groupConfigMap[gGrpId].withParamModel_9 = 1;
        groupConfigMap[gGrpId].isLIF = 0;
        KERNEL_INFO("Set a nine parameter group!");
    }
}
 
 
// set LIF parameters for the group
void SNN::setNeuronParametersLIF(int gGrpId, int tau_m, int tau_ref, float vTh, float vReset, double minRmem, double maxRmem)
{
    assert(gGrpId >= -1);
    assert(tau_m >= 0); assert(tau_ref >= 0); assert(vReset < vTh);
    assert(minRmem >= 0.0f); assert(minRmem <= maxRmem);
 
    if (gGrpId == ALL) { // shortcut for all groups
        for(int grpId = 0; grpId < numGroups; grpId++) {
            setNeuronParametersLIF(grpId, tau_m, tau_ref, vTh, vReset, minRmem, maxRmem);
        }
    } else {
        groupConfigMap[gGrpId].neuralDynamicsConfig.lif_tau_m = tau_m;
        groupConfigMap[gGrpId].neuralDynamicsConfig.lif_tau_ref = tau_ref;
        groupConfigMap[gGrpId].neuralDynamicsConfig.lif_vTh = vTh;
        groupConfigMap[gGrpId].neuralDynamicsConfig.lif_vReset = vReset;
        groupConfigMap[gGrpId].neuralDynamicsConfig.lif_minRmem = minRmem;
        groupConfigMap[gGrpId].neuralDynamicsConfig.lif_maxRmem = maxRmem;
        groupConfigMap[gGrpId].withParamModel_9 = 0;
        groupConfigMap[gGrpId].isLIF = 1;
    }
}
 
void SNN::setNeuromodulator(int gGrpId, float baseDP, float tauDP, float base5HT, float tau5HT, float baseACh,
    float tauACh, float baseNE, float tauNE) {
 
    assert(gGrpId >= -1);
    assert(baseDP > 0.0f); assert(base5HT > 0.0f); assert(baseACh > 0.0f); assert(baseNE > 0.0f);
    assert(tauDP > 0); assert(tau5HT > 0); assert(tauACh > 0); assert(tauNE > 0);
 
    if (gGrpId == ALL) { // shortcut for all groups
        for (int grpId = 0; grpId < numGroups; grpId++) {
            setNeuromodulator(grpId, baseDP, tauDP, base5HT, tau5HT, baseACh, tauACh, baseNE, tauNE);
        }
    } else {
        groupConfigMap[gGrpId].neuromodulatorConfig.baseDP = baseDP;
        groupConfigMap[gGrpId].neuromodulatorConfig.decayDP = 1.0f - (1.0f / tauDP);
        groupConfigMap[gGrpId].neuromodulatorConfig.base5HT = base5HT;
        groupConfigMap[gGrpId].neuromodulatorConfig.decay5HT = 1.0f - (1.0f / tau5HT);
        groupConfigMap[gGrpId].neuromodulatorConfig.baseACh = baseACh;
        groupConfigMap[gGrpId].neuromodulatorConfig.decayACh = 1.0f - (1.0f / tauACh);
        groupConfigMap[gGrpId].neuromodulatorConfig.baseNE = baseNE;
        groupConfigMap[gGrpId].neuromodulatorConfig.decayNE = 1.0f - (1.0f / tauNE);
    }
}
 
// set ESTDP params
void SNN::setESTDP(int gGrpId, bool isSet, STDPType type, STDPCurve curve, float alphaPlus, float tauPlus, float alphaMinus, float tauMinus, float gamma) {
    assert(gGrpId >= -1);
    if (isSet) {
        assert(type!=UNKNOWN_STDP);
        assert(tauPlus > 0.0f); assert(tauMinus > 0.0f); assert(gamma >= 0.0f);
    }
 
    if (gGrpId == ALL) { // shortcut for all groups
        for(int grpId = 0; grpId < numGroups; grpId++) {
            setESTDP(grpId, isSet, type, curve, alphaPlus, tauPlus, alphaMinus, tauMinus, gamma);
        }
    } else {
        // set STDP for a given group
        // set params for STDP curve
        groupConfigMap[gGrpId].stdpConfig.ALPHA_PLUS_EXC    = alphaPlus;
        groupConfigMap[gGrpId].stdpConfig.ALPHA_MINUS_EXC   = alphaMinus;
        groupConfigMap[gGrpId].stdpConfig.TAU_PLUS_INV_EXC  = 1.0f / tauPlus;
        groupConfigMap[gGrpId].stdpConfig.TAU_MINUS_INV_EXC = 1.0f / tauMinus;
        groupConfigMap[gGrpId].stdpConfig.GAMMA             = gamma;
        groupConfigMap[gGrpId].stdpConfig.KAPPA             = (1 + exp(-gamma / tauPlus)) / (1 - exp(-gamma / tauPlus));
        groupConfigMap[gGrpId].stdpConfig.OMEGA             = alphaPlus * (1 - groupConfigMap[gGrpId].stdpConfig.KAPPA);
        // set flags for STDP function
        groupConfigMap[gGrpId].stdpConfig.WithESTDPtype = type;
        groupConfigMap[gGrpId].stdpConfig.WithESTDPcurve = curve;
        groupConfigMap[gGrpId].stdpConfig.WithESTDP     = isSet;
        groupConfigMap[gGrpId].stdpConfig.WithSTDP      |= groupConfigMap[gGrpId].stdpConfig.WithESTDP;
        sim_with_stdp                                   |= groupConfigMap[gGrpId].stdpConfig.WithSTDP;
 
        KERNEL_INFO("E-STDP %s for %s(%d)", isSet?"enabled":"disabled", groupConfigMap[gGrpId].grpName.c_str(), gGrpId);
    }
}
 
// set ISTDP params
void SNN::setISTDP(int gGrpId, bool isSet, STDPType type, STDPCurve curve, float ab1, float ab2, float tau1, float tau2) {
    assert(gGrpId >= -1);
    if (isSet) {
        assert(type != UNKNOWN_STDP);
        assert(tau1 > 0); assert(tau2 > 0);
    }
 
    if (gGrpId==ALL) { // shortcut for all groups
        for(int grpId = 0; grpId < numGroups; grpId++) {
            setISTDP(grpId, isSet, type, curve, ab1, ab2, tau1, tau2);
        }
    } else {
        // set STDP for a given group
        // set params for STDP curve
        if (curve == EXP_CURVE) {
            groupConfigMap[gGrpId].stdpConfig.ALPHA_PLUS_INB = ab1;
            groupConfigMap[gGrpId].stdpConfig.ALPHA_MINUS_INB = ab2;
            groupConfigMap[gGrpId].stdpConfig.TAU_PLUS_INV_INB = 1.0f / tau1;
            groupConfigMap[gGrpId].stdpConfig.TAU_MINUS_INV_INB = 1.0f / tau2;
            groupConfigMap[gGrpId].stdpConfig.BETA_LTP      = 0.0f;
            groupConfigMap[gGrpId].stdpConfig.BETA_LTD      = 0.0f;
            groupConfigMap[gGrpId].stdpConfig.LAMBDA            = 1.0f;
            groupConfigMap[gGrpId].stdpConfig.DELTA         = 1.0f;
        } else {
            groupConfigMap[gGrpId].stdpConfig.ALPHA_PLUS_INB = 0.0f;
            groupConfigMap[gGrpId].stdpConfig.ALPHA_MINUS_INB = 0.0f;
            groupConfigMap[gGrpId].stdpConfig.TAU_PLUS_INV_INB = 1.0f;
            groupConfigMap[gGrpId].stdpConfig.TAU_MINUS_INV_INB = 1.0f;
            groupConfigMap[gGrpId].stdpConfig.BETA_LTP      = ab1;
            groupConfigMap[gGrpId].stdpConfig.BETA_LTD      = ab2;
            groupConfigMap[gGrpId].stdpConfig.LAMBDA            = tau1;
            groupConfigMap[gGrpId].stdpConfig.DELTA         = tau2;
        }
        // set flags for STDP function
        //FIXME: separate STDPType to ESTDPType and ISTDPType
        groupConfigMap[gGrpId].stdpConfig.WithISTDPtype = type;
        groupConfigMap[gGrpId].stdpConfig.WithISTDPcurve = curve;
        groupConfigMap[gGrpId].stdpConfig.WithISTDP     = isSet;
        groupConfigMap[gGrpId].stdpConfig.WithSTDP      |= groupConfigMap[gGrpId].stdpConfig.WithISTDP;
        sim_with_stdp                   |= groupConfigMap[gGrpId].stdpConfig.WithSTDP;
 
        KERNEL_INFO("I-STDP %s for %s(%d)", isSet?"enabled":"disabled", groupConfigMap[gGrpId].grpName.c_str(), gGrpId);
    }
}
 
// set STP params
void SNN::setSTP(int gGrpId, bool isSet, float STP_U, float STP_tau_u, float STP_tau_x) {
    assert(gGrpId >= -1);
    if (isSet) {
        assert(STP_U > 0 && STP_U <= 1); assert(STP_tau_u > 0); assert(STP_tau_x > 0);
    }
 
    if (gGrpId == ALL) { // shortcut for all groups
        for(int grpId = 0; grpId < numGroups; grpId++) {
            setSTP(grpId, isSet, STP_U, STP_tau_u, STP_tau_x);
        }
    } else {
        // set STDP for a given group
        sim_with_stp                                    |= isSet;
        groupConfigMap[gGrpId].stpConfig.WithSTP        = isSet;
        groupConfigMap[gGrpId].stpConfig.STP_A          = (STP_U > 0.0f) ? 1.0 / STP_U : 1.0f; // scaling factor
        groupConfigMap[gGrpId].stpConfig.STP_U          = STP_U;
        groupConfigMap[gGrpId].stpConfig.STP_tau_u_inv  = 1.0f / STP_tau_u; // facilitatory
        groupConfigMap[gGrpId].stpConfig.STP_tau_x_inv  = 1.0f / STP_tau_x; // depressive
 
        KERNEL_INFO("STP %s for %d (%s):\tA: %1.4f, U: %1.4f, tau_u: %4.0f, tau_x: %4.0f", isSet?"enabled":"disabled",
                    gGrpId, groupConfigMap[gGrpId].grpName.c_str(), groupConfigMap[gGrpId].stpConfig.STP_A, STP_U, STP_tau_u, STP_tau_x);
    }
}
 
void SNN::setWeightAndWeightChangeUpdate(UpdateInterval wtANDwtChangeUpdateInterval, bool enableWtChangeDecay, float wtChangeDecay) {
    assert(wtChangeDecay > 0.0f && wtChangeDecay < 1.0f);
 
    switch (wtANDwtChangeUpdateInterval) {
        case INTERVAL_10MS:
            wtANDwtChangeUpdateInterval_ = 10;
            break;
        case INTERVAL_100MS:
            wtANDwtChangeUpdateInterval_ = 100;
            break;
        case INTERVAL_1000MS:
        default:
            wtANDwtChangeUpdateInterval_ = 1000;
            break;
    }
 
    if (enableWtChangeDecay) {
        // set up stdp factor according to update interval
        switch (wtANDwtChangeUpdateInterval) {
        case INTERVAL_10MS:
            stdpScaleFactor_ = 0.005f;
            break;
        case INTERVAL_100MS:
            stdpScaleFactor_ = 0.05f;
            break;
        case INTERVAL_1000MS:
        default:
            stdpScaleFactor_ = 0.5f;
            break;
        }
        // set up weight decay
        wtChangeDecay_ = wtChangeDecay;
    } else {
        stdpScaleFactor_ = 1.0f;
        wtChangeDecay_ = 0.0f;
    }
 
    KERNEL_INFO("Update weight and weight change every %d ms", wtANDwtChangeUpdateInterval_);
    KERNEL_INFO("Weight Change Decay is %s", enableWtChangeDecay? "enabled" : "disable");
    KERNEL_INFO("STDP scale factor = %1.3f, wtChangeDecay = %1.3f", stdpScaleFactor_, wtChangeDecay_);
}
 
 
// reorganize the network and do the necessary allocation
// of all variable for carrying out the simulation..
// this code is run only one time during network initialization
void SNN::setupNetwork() {
    switch (snnState) {
    case CONFIG_SNN:
        compileSNN();
    case COMPILED_SNN:
        partitionSNN();
    case PARTITIONED_SNN:
        generateRuntimeSNN();
        break;
    case EXECUTABLE_SNN:
        break;
    default:
        KERNEL_ERROR("Unknown SNN state");
        break;
    }
}
 
 
int SNN::runNetwork(int _nsec, int _nmsec, bool printRunSummary) {
    assert(_nmsec >= 0 && _nmsec < 1000);
    assert(_nsec  >= 0);
    int runDurationMs = _nsec*1000 + _nmsec;
    KERNEL_DEBUG("runNetwork: runDur=%dms, printRunSummary=%s", runDurationMs, printRunSummary?"y":"n");
 
    // setupNetwork() must have already been called
    assert(snnState == EXECUTABLE_SNN);
 
    // don't bother printing if logger mode is SILENT
    printRunSummary = (loggerMode_==SILENT) ? false : printRunSummary;
 
    // first-time run: inform the user the simulation is running now
    if (simTime==0 && printRunSummary) {
        KERNEL_INFO("");
        KERNEL_INFO("******************** Running the simulation on %d GPU(s) and %d CPU(s) ***************************", numGPUs, numCores);
        KERNEL_INFO("");
    }
 
    // reset all spike counters
    resetSpikeCnt(ALL);
 
    // store current start time for future reference
    simTimeRunStart = simTime;
    simTimeRunStop  = simTime + runDurationMs;
    assert(simTimeRunStop >= simTimeRunStart); // check for arithmetic underflow
 
    // ConnectionMonitor is a special case: we might want the first snapshot at t=0 in the binary
    // but updateTime() is false for simTime==0.
    // And we cannot put this code in ConnectionMonitorCore::init, because then the user would have no
    // way to call ConnectionMonitor::setUpdateTimeIntervalSec before...
    if (simTime == 0 && numConnectionMonitor) {
        updateConnectionMonitor();
    }
 
    // set the Poisson generation time slice to be at the run duration up to MAX_TIME_SLICE
    setGrpTimeSlice(ALL, std::max(1, std::min(runDurationMs, MAX_TIME_SLICE)));
 
#ifndef __NO_CUDA__
    CUDA_RESET_TIMER(timer);
    CUDA_START_TIMER(timer);
#endif
 
    //KERNEL_INFO("Reached the advSimStep loop!");
 
    // if nsec=0, simTimeMs=10, we need to run the simulator for 10 timeStep;
    // if nsec=1, simTimeMs=10, we need to run the simulator for 1*1000+10, time Step;
    for(int i = 0; i < runDurationMs; i++) {
        advSimStep();
        //KERNEL_INFO("Executed an advSimStep!");
 
        // update weight every updateInterval ms if plastic synapses present
        if (!sim_with_fixedwts && wtANDwtChangeUpdateInterval_ == ++wtANDwtChangeUpdateIntervalCnt_) {
            wtANDwtChangeUpdateIntervalCnt_ = 0; // reset counter
            if (!sim_in_testing) {
                // keep this if statement separate from the above, so that the counter is updated correctly
                updateWeights();
            }
        }
 
        // Note: updateTime() advance simTime, simTimeMs, and simTimeSec accordingly
        if (updateTime()) {
            // finished one sec of simulation...
            if (numSpikeMonitor) {
                updateSpikeMonitor();
            }
            if (numGroupMonitor) {
                updateGroupMonitor();
            }
            if (numConnectionMonitor) {
                updateConnectionMonitor();
            }
            if (numNeuronMonitor) {
                updateNeuronMonitor();
            }
 
            shiftSpikeTables();
        }
 
        fetchNeuronSpikeCount(ALL);
    }
 
    //KERNEL_INFO("Updated monitors!");
 
    // user can opt to display some runNetwork summary
    if (printRunSummary) {
 
        // if there are Monitors available and it's time to show the log, print status for each group
        if (numSpikeMonitor) {
            printStatusSpikeMonitor(ALL);
        }
        if (numConnectionMonitor) {
            printStatusConnectionMonitor(ALL);
        }
        if (numGroupMonitor) {
            printStatusGroupMonitor(ALL);
        }
 
        // record time of run summary print
        simTimeLastRunSummary = simTime;
    }
 
    // call updateSpike(Group)Monitor again to fetch all the left-over spikes and group status (neuromodulator)
    updateSpikeMonitor();
    updateGroupMonitor();
 
    // keep track of simulation time...
#ifndef __NO_CUDA__
    CUDA_STOP_TIMER(timer);
    lastExecutionTime = CUDA_GET_TIMER_VALUE(timer);
    cumExecutionTime += lastExecutionTime;
#endif
    return 0;
}
 
 
 
 
// adds a bias to every weight in the connection
void SNN::biasWeights(short int connId, float bias, bool updateWeightRange) {
    assert(connId>=0 && connId<numConnections);
 
    int netId = groupConfigMDMap[connectConfigMap[connId].grpDest].netId;
    int lGrpId = groupConfigMDMap[connectConfigMap[connId].grpDest].lGrpId;
 
    fetchPreConnectionInfo(netId);
    fetchConnIdsLookupArray(netId);
    fetchSynapseState(netId);
    // iterate over all postsynaptic neurons
    for (int lNId = groupConfigs[netId][lGrpId].lStartN; lNId <= groupConfigs[netId][lGrpId].lEndN; lNId++) {
        unsigned int cumIdx = managerRuntimeData.cumulativePre[lNId];
 
        // iterate over all presynaptic neurons
        unsigned int pos_ij = cumIdx;
        for (int j = 0; j < managerRuntimeData.Npre[lNId]; pos_ij++, j++) {
            if (managerRuntimeData.connIdsPreIdx[pos_ij] == connId) {
                // apply bias to weight
                float weight = managerRuntimeData.wt[pos_ij] + bias;
 
                // inform user of acton taken if weight is out of bounds
//              bool needToPrintDebug = (weight+bias>connInfo->maxWt || weight+bias<connInfo->minWt);
                bool needToPrintDebug = (weight > connectConfigMap[connId].maxWt || weight < 0.0f);
 
                if (updateWeightRange) {
                    // if this flag is set, we need to update minWt,maxWt accordingly
                    // will be saving new maxSynWt and copying to GPU below
//                  connInfo->minWt = fmin(connInfo->minWt, weight);
                    connectConfigMap[connId].maxWt = std::max(connectConfigMap[connId].maxWt, weight);
                    if (needToPrintDebug) {
                        KERNEL_DEBUG("biasWeights(%d,%f,%s): updated weight ranges to [%f,%f]", connId, bias,
                            (updateWeightRange?"true":"false"), 0.0f, connectConfigMap[connId].maxWt);
                    }
                } else {
                    // constrain weight to boundary values
                    // compared to above, we swap minWt/maxWt logic
                    weight = std::min(weight, connectConfigMap[connId].maxWt);
//                  weight = fmax(weight, connInfo->minWt);
                    weight = std::max(weight, 0.0f);
                    if (needToPrintDebug) {
                        KERNEL_DEBUG("biasWeights(%d,%f,%s): constrained weight %f to [%f,%f]", connId, bias,
                            (updateWeightRange?"true":"false"), weight, 0.0f, connectConfigMap[connId].maxWt);
                    }
                }
 
                // update datastructures
                managerRuntimeData.wt[pos_ij] = weight;
                managerRuntimeData.maxSynWt[pos_ij] = connectConfigMap[connId].maxWt; // it's easier to just update, even if it hasn't changed
            }
        }
 
        // update GPU datastructures in batches, grouped by post-neuron
        if (netId < CPU_RUNTIME_BASE) {
#ifndef __NO_CUDA__
            CUDA_CHECK_ERRORS( cudaMemcpy(&(runtimeData[netId].wt[cumIdx]), &(managerRuntimeData.wt[cumIdx]), sizeof(float)*managerRuntimeData.Npre[lNId],
                cudaMemcpyHostToDevice) );
 
            if (runtimeData[netId].maxSynWt != NULL) {
                // only copy maxSynWt if datastructure actually exists on the GPU runtime
                // (that logic should be done elsewhere though)
                CUDA_CHECK_ERRORS( cudaMemcpy(&(runtimeData[netId].maxSynWt[cumIdx]), &(managerRuntimeData.maxSynWt[cumIdx]),
                    sizeof(float) * managerRuntimeData.Npre[lNId], cudaMemcpyHostToDevice) );
            }
#else
            assert(false);
#endif
        } else {
            memcpy(&runtimeData[netId].wt[cumIdx], &managerRuntimeData.wt[cumIdx], sizeof(float) * managerRuntimeData.Npre[lNId]);
 
            if (runtimeData[netId].maxSynWt != NULL) {
                // only copy maxSynWt if datastructure actually exists on the CPU runtime
                // (that logic should be done elsewhere though)
                memcpy(&runtimeData[netId].maxSynWt[cumIdx], &managerRuntimeData.maxSynWt[cumIdx], sizeof(float) * managerRuntimeData.Npre[lNId]);
            }
        }
    }
}
 
// deallocates dynamical structures and exits
void SNN::exitSimulation(int val) {
    deleteObjects();
    exit(val);
}
 
// reads network state from file
void SNN::loadSimulation(FILE* fid) {
    loadSimFID = fid;
}
 
// multiplies every weight with a scaling factor
void SNN::scaleWeights(short int connId, float scale, bool updateWeightRange) {
    assert(connId>=0 && connId<numConnections);
    assert(scale>=0.0f);
 
    int netId = groupConfigMDMap[connectConfigMap[connId].grpDest].netId;
    int lGrpId = groupConfigMDMap[connectConfigMap[connId].grpDest].lGrpId;
 
    fetchPreConnectionInfo(netId);
    fetchConnIdsLookupArray(netId);
    fetchSynapseState(netId);
 
    // iterate over all postsynaptic neurons
    for (int lNId = groupConfigs[netId][lGrpId].lStartN; lNId <= groupConfigs[netId][lGrpId].lEndN; lNId++) {
        unsigned int cumIdx = managerRuntimeData.cumulativePre[lNId];
 
        // iterate over all presynaptic neurons
        unsigned int pos_ij = cumIdx;
        for (int j = 0; j < managerRuntimeData.Npre[lNId]; pos_ij++, j++) {
            if (managerRuntimeData.connIdsPreIdx[pos_ij]==connId) {
                // apply bias to weight
                float weight = managerRuntimeData.wt[pos_ij] * scale;
 
                // inform user of acton taken if weight is out of bounds
//              bool needToPrintDebug = (weight>connInfo->maxWt || weight<connInfo->minWt);
                bool needToPrintDebug = (weight > connectConfigMap[connId].maxWt || weight < 0.0f);
 
                if (updateWeightRange) {
                    // if this flag is set, we need to update minWt,maxWt accordingly
                    // will be saving new maxSynWt and copying to GPU below
//                  connInfo->minWt = fmin(connInfo->minWt, weight);
                    connectConfigMap[connId].maxWt = std::max(connectConfigMap[connId].maxWt, weight);
                    if (needToPrintDebug) {
                        KERNEL_DEBUG("scaleWeights(%d,%f,%s): updated weight ranges to [%f,%f]", connId, scale,
                            (updateWeightRange?"true":"false"), 0.0f, connectConfigMap[connId].maxWt);
                    }
                } else {
                    // constrain weight to boundary values
                    // compared to above, we swap minWt/maxWt logic
                    weight = std::min(weight, connectConfigMap[connId].maxWt);
//                  weight = fmax(weight, connInfo->minWt);
                    weight = std::max(weight, 0.0f);
                    if (needToPrintDebug) {
                        KERNEL_DEBUG("scaleWeights(%d,%f,%s): constrained weight %f to [%f,%f]", connId, scale,
                            (updateWeightRange?"true":"false"), weight, 0.0f, connectConfigMap[connId].maxWt);
                    }
                }
 
                // update datastructures
                managerRuntimeData.wt[pos_ij] = weight;
                managerRuntimeData.maxSynWt[pos_ij] = connectConfigMap[connId].maxWt; // it's easier to just update, even if it hasn't changed
            }
        }
 
        // update GPU datastructures in batches, grouped by post-neuron
        if (netId < CPU_RUNTIME_BASE) {
#ifndef __NO_CUDA__
            CUDA_CHECK_ERRORS(cudaMemcpy(&runtimeData[netId].wt[cumIdx], &managerRuntimeData.wt[cumIdx], sizeof(float)*managerRuntimeData.Npre[lNId],
                cudaMemcpyHostToDevice));
 
            if (runtimeData[netId].maxSynWt != NULL) {
                // only copy maxSynWt if datastructure actually exists on the GPU runtime
                // (that logic should be done elsewhere though)
                CUDA_CHECK_ERRORS(cudaMemcpy(&runtimeData[netId].maxSynWt[cumIdx], &managerRuntimeData.maxSynWt[cumIdx],
                    sizeof(float) * managerRuntimeData.Npre[lNId], cudaMemcpyHostToDevice));
            }
#else
            assert(false);
#endif
        } else {
            memcpy(&runtimeData[netId].wt[cumIdx], &managerRuntimeData.wt[cumIdx], sizeof(float) * managerRuntimeData.Npre[lNId]);
 
            if (runtimeData[netId].maxSynWt != NULL) {
                // only copy maxSynWt if datastructure actually exists on the CPU runtime
                // (that logic should be done elsewhere though)
                memcpy(&runtimeData[netId].maxSynWt[cumIdx], &managerRuntimeData.maxSynWt[cumIdx], sizeof(float) * managerRuntimeData.Npre[lNId]);
            }
        }
    }
}
 
// FIXME: distinguish the function call at CONFIG_STATE and SETUP_STATE, where groupConfigs[0][] might not be available
// or groupConfigMap is not sync with groupConfigs[0][]
GroupMonitor* SNN::setGroupMonitor(int gGrpId, FILE* fid) {
    int netId = groupConfigMDMap[gGrpId].netId;
    int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
 
    // check whether group already has a GroupMonitor
    if (groupConfigMDMap[gGrpId].groupMonitorId >= 0) {
        KERNEL_ERROR("setGroupMonitor has already been called on Group %d (%s).", gGrpId, groupConfigMap[gGrpId].grpName.c_str());
        exitSimulation(1);
    }
 
    // create new GroupMonitorCore object in any case and initialize analysis components
    // grpMonObj destructor (see below) will deallocate it
    GroupMonitorCore* grpMonCoreObj = new GroupMonitorCore(this, numGroupMonitor, gGrpId);
    groupMonCoreList[numGroupMonitor] = grpMonCoreObj;
 
    // assign group status file ID if we selected to write to a file, else it's NULL
    // if file pointer exists, it has already been fopened
    // this will also write the header section of the group status file
    // grpMonCoreObj destructor will fclose it
    grpMonCoreObj->setGroupFileId(fid);
 
    // create a new GroupMonitor object for the user-interface
    // SNN::deleteObjects will deallocate it
    GroupMonitor* grpMonObj = new GroupMonitor(grpMonCoreObj);
    groupMonList[numGroupMonitor] = grpMonObj;
 
    // also inform the group that it is being monitored...
    groupConfigMDMap[gGrpId].groupMonitorId = numGroupMonitor;
 
    numGroupMonitor++;
    KERNEL_INFO("GroupMonitor set for group %d (%s)", gGrpId, groupConfigMap[gGrpId].grpName.c_str());
 
    return grpMonObj;
}
 
// FIXME: distinguish the function call at CONFIG_STATE and SETUP_STATE, where group(connect)Config[] might not be available
// or group(connect)ConfigMap is not sync with group(connect)Config[]
ConnectionMonitor* SNN::setConnectionMonitor(int grpIdPre, int grpIdPost, FILE* fid) {
    // find connection based on pre-post pair
    short int connId = getConnectId(grpIdPre, grpIdPost);
    if (connId<0) {
        KERNEL_ERROR("No connection found from group %d(%s) to group %d(%s)", grpIdPre, getGroupName(grpIdPre).c_str(),
            grpIdPost, getGroupName(grpIdPost).c_str());
        exitSimulation(1);
    }
 
    // check whether connection already has a connection monitor
    if (connectConfigMap[connId].connectionMonitorId >= 0) {
        KERNEL_ERROR("setConnectionMonitor has already been called on Connection %d (MonitorId=%d)", connId, connectConfigMap[connId].connectionMonitorId);
        exitSimulation(1);
    }
 
    // inform the connection that it is being monitored...
    // this needs to be called before new ConnectionMonitorCore
    connectConfigMap[connId].connectionMonitorId = numConnectionMonitor;
 
    // create new ConnectionMonitorCore object in any case and initialize
    // connMonObj destructor (see below) will deallocate it
    ConnectionMonitorCore* connMonCoreObj = new ConnectionMonitorCore(this, numConnectionMonitor, connId,
        grpIdPre, grpIdPost);
    connMonCoreList[numConnectionMonitor] = connMonCoreObj;
 
    // assign conn file ID if we selected to write to a file, else it's NULL
    // if file pointer exists, it has already been fopened
    // this will also write the header section of the conn file
    // connMonCoreObj destructor will fclose it
    connMonCoreObj->setConnectFileId(fid);
 
    // create a new ConnectionMonitor object for the user-interface
    // SNN::deleteObjects will deallocate it
    ConnectionMonitor* connMonObj = new ConnectionMonitor(connMonCoreObj);
    connMonList[numConnectionMonitor] = connMonObj;
 
    // now init core object (depends on several datastructures allocated above)
    connMonCoreObj->init();
 
    numConnectionMonitor++;
    KERNEL_INFO("ConnectionMonitor %d set for Connection %d: %d(%s) => %d(%s)", connectConfigMap[connId].connectionMonitorId, connId, grpIdPre, getGroupName(grpIdPre).c_str(),
        grpIdPost, getGroupName(grpIdPost).c_str());
 
    return connMonObj;
}
 
// FIXME: distinguish the function call at CONFIG_STATE and SETUP_STATE, where groupConfigs[0][] might not be available
// or groupConfigMap is not sync with groupConfigs[0][]
// sets up a spike generator
void SNN::setSpikeGenerator(int gGrpId, SpikeGeneratorCore* spikeGenFunc) {
    assert(snnState == CONFIG_SNN); // must be called before setupNetwork() to work on GPU
    assert(spikeGenFunc);
    assert(groupConfigMap[gGrpId].isSpikeGenerator);
    groupConfigMap[gGrpId].spikeGenFunc = spikeGenFunc;
}
 
// record spike information, return a SpikeInfo object
SpikeMonitor* SNN::setSpikeMonitor(int gGrpId, FILE* fid) {
    // check whether group already has a SpikeMonitor
    if (groupConfigMDMap[gGrpId].spikeMonitorId >= 0) {
        // in this case, return the current object and update fid
        SpikeMonitor* spkMonObj = getSpikeMonitor(gGrpId);
 
        // update spike file ID
        SpikeMonitorCore* spkMonCoreObj = getSpikeMonitorCore(gGrpId);
        spkMonCoreObj->setSpikeFileId(fid);
 
        KERNEL_INFO("SpikeMonitor updated for group %d (%s)", gGrpId, groupConfigMap[gGrpId].grpName.c_str());
        return spkMonObj;
    } else {
        // create new SpikeMonitorCore object in any case and initialize analysis components
        // spkMonObj destructor (see below) will deallocate it
        SpikeMonitorCore* spkMonCoreObj = new SpikeMonitorCore(this, numSpikeMonitor, gGrpId);
        spikeMonCoreList[numSpikeMonitor] = spkMonCoreObj;
 
        // assign spike file ID if we selected to write to a file, else it's NULL
        // if file pointer exists, it has already been fopened
        // this will also write the header section of the spike file
        // spkMonCoreObj destructor will fclose it
        spkMonCoreObj->setSpikeFileId(fid);
 
        // create a new SpikeMonitor object for the user-interface
        // SNN::deleteObjects will deallocate it
        SpikeMonitor* spkMonObj = new SpikeMonitor(spkMonCoreObj);
        spikeMonList[numSpikeMonitor] = spkMonObj;
 
        // also inform the grp that it is being monitored...
        groupConfigMDMap[gGrpId].spikeMonitorId = numSpikeMonitor;
 
        numSpikeMonitor++;
        KERNEL_INFO("SpikeMonitor set for group %d (%s)", gGrpId, groupConfigMap[gGrpId].grpName.c_str());
 
        return spkMonObj;
    }
}
 
// record neuron state information, return a NeuronInfo object
NeuronMonitor* SNN::setNeuronMonitor(int gGrpId, FILE* fid) {
    int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
    int netId = groupConfigMDMap[gGrpId].netId;
 
    if (getGroupNumNeurons(gGrpId) > 128) {
        KERNEL_WARN("Due to limited memory space, only the first 128 neurons can be monitored by NeuronMonitor");
    }
 
    // check whether group already has a NeuronMonitor
    if (groupConfigMDMap[gGrpId].neuronMonitorId >= 0) {
        // in this case, return the current object and update fid
        NeuronMonitor* nrnMonObj = getNeuronMonitor(gGrpId);
 
        // update neuron file ID
        NeuronMonitorCore* nrnMonCoreObj = getNeuronMonitorCore(gGrpId);
        nrnMonCoreObj->setNeuronFileId(fid);
 
        KERNEL_INFO("NeuronMonitor updated for group %d (%s)", gGrpId, groupConfigMap[gGrpId].grpName.c_str());
        return nrnMonObj;
    } else {
        // create new NeuronMonitorCore object in any case and initialize analysis components
        // nrnMonObj destructor (see below) will deallocate it
        NeuronMonitorCore* nrnMonCoreObj = new NeuronMonitorCore(this, numNeuronMonitor, gGrpId);
        neuronMonCoreList[numNeuronMonitor] = nrnMonCoreObj;
 
        // assign neuron state file ID if we selected to write to a file, else it's NULL
        // if file pointer exists, it has already been fopened
        // this will also write the header section of the spike file
        // spkMonCoreObj destructor will fclose it
        nrnMonCoreObj->setNeuronFileId(fid);
 
        // create a new NeuronMonitor object for the user-interface
        // SNN::deleteObjects will deallocate it
        NeuronMonitor* nrnMonObj = new NeuronMonitor(nrnMonCoreObj);
        neuronMonList[numNeuronMonitor] = nrnMonObj;
 
        // also inform the grp that it is being monitored...
        groupConfigMDMap[gGrpId].neuronMonitorId = numNeuronMonitor;
 
        numNeuronMonitor++;
        KERNEL_INFO("NeuronMonitor set for group %d (%s)", gGrpId, groupConfigMap[gGrpId].grpName.c_str());
 
        return nrnMonObj;
    }
}
 
// FIXME: distinguish the function call at CONFIG_STATE and RUN_STATE, where groupConfigs[0][] might not be available
// or groupConfigMap is not sync with groupConfigs[0][]
// assigns spike rate to group
void SNN::setSpikeRate(int gGrpId, PoissonRate* ratePtr, int refPeriod) {
    int netId = groupConfigMDMap[gGrpId].netId;
    int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
 
    assert(gGrpId >= 0 && lGrpId < networkConfigs[netId].numGroups);
    assert(ratePtr);
    assert(groupConfigMap[gGrpId].isSpikeGenerator);
    assert(ratePtr->getNumNeurons() == groupConfigMap[gGrpId].numN);
    assert(refPeriod >= 1);
 
    groupConfigMDMap[gGrpId].ratePtr = ratePtr;
    groupConfigMDMap[gGrpId].refractPeriod = refPeriod;
    spikeRateUpdated = true;
}
 
// sets the weight value of a specific synapse
void SNN::setWeight(short int connId, int neurIdPre, int neurIdPost, float weight, bool updateWeightRange) {
    assert(connId>=0 && connId<getNumConnections());
    assert(weight>=0.0f);
 
    assert(neurIdPre >= 0  && neurIdPre < getGroupNumNeurons(connectConfigMap[connId].grpSrc));
    assert(neurIdPost >= 0 && neurIdPost < getGroupNumNeurons(connectConfigMap[connId].grpDest));
 
    float maxWt = fabs(connectConfigMap[connId].maxWt);
    float minWt = 0.0f;
 
    // inform user of acton taken if weight is out of bounds
    bool needToPrintDebug = (weight>maxWt || weight<minWt);
 
    int netId = groupConfigMDMap[connectConfigMap[connId].grpDest].netId;
    int postlGrpId = groupConfigMDMap[connectConfigMap[connId].grpDest].lGrpId;
    int prelGrpId = groupConfigMDMap[connectConfigMap[connId].grpSrc].lGrpId;
 
    fetchPreConnectionInfo(netId);
    fetchConnIdsLookupArray(netId);
    fetchSynapseState(netId);
 
    if (updateWeightRange) {
        // if this flag is set, we need to update minWt,maxWt accordingly
        // will be saving new maxSynWt and copying to GPU below
//      connInfo->minWt = fmin(connInfo->minWt, weight);
        maxWt = fmax(maxWt, weight);
        if (needToPrintDebug) {
            KERNEL_DEBUG("setWeight(%d,%d,%d,%f,%s): updated weight ranges to [%f,%f]", connId, neurIdPre, neurIdPost,
                weight, (updateWeightRange?"true":"false"), minWt, maxWt);
        }
    } else {
        // constrain weight to boundary values
        // compared to above, we swap minWt/maxWt logic
        weight = fmin(weight, maxWt);
        weight = fmax(weight, minWt);
        if (needToPrintDebug) {
            KERNEL_DEBUG("setWeight(%d,%d,%d,%f,%s): constrained weight %f to [%f,%f]", connId, neurIdPre, neurIdPost,
                weight, (updateWeightRange?"true":"false"), weight, minWt, maxWt);
        }
    }
 
    // find real ID of pre- and post-neuron
    int neurIdPreReal = groupConfigs[netId][prelGrpId].lStartN + neurIdPre;
    int neurIdPostReal = groupConfigs[netId][postlGrpId].lStartN + neurIdPost;
 
    // iterate over all presynaptic synapses until right one is found
    bool synapseFound = false;
    int pos_ij = managerRuntimeData.cumulativePre[neurIdPostReal];
    for (int j = 0; j < managerRuntimeData.Npre[neurIdPostReal]; pos_ij++, j++) {
        SynInfo* preId = &(managerRuntimeData.preSynapticIds[pos_ij]);
        int pre_nid = GET_CONN_NEURON_ID((*preId));
        if (GET_CONN_NEURON_ID((*preId)) == neurIdPreReal) {
            assert(managerRuntimeData.connIdsPreIdx[pos_ij] == connId); // make sure we've got the right connection ID
 
            managerRuntimeData.wt[pos_ij] = isExcitatoryGroup(connectConfigMap[connId].grpSrc) ? weight : -1.0 * weight;
            managerRuntimeData.maxSynWt[pos_ij] = isExcitatoryGroup(connectConfigMap[connId].grpSrc) ? maxWt : -1.0 * maxWt;
 
            if (netId < CPU_RUNTIME_BASE) {
#ifndef __NO_CUDA__
                // need to update datastructures on GPU runtime
                CUDA_CHECK_ERRORS(cudaMemcpy(&runtimeData[netId].wt[pos_ij], &managerRuntimeData.wt[pos_ij], sizeof(float), cudaMemcpyHostToDevice));
                if (runtimeData[netId].maxSynWt != NULL) {
                    // only copy maxSynWt if datastructure actually exists on the GPU runtime
                    // (that logic should be done elsewhere though)
                    CUDA_CHECK_ERRORS(cudaMemcpy(&runtimeData[netId].maxSynWt[pos_ij], &managerRuntimeData.maxSynWt[pos_ij], sizeof(float), cudaMemcpyHostToDevice));
                }
#else
                assert(false);
#endif
            } else {
                // need to update datastructures on CPU runtime
                memcpy(&runtimeData[netId].wt[pos_ij], &managerRuntimeData.wt[pos_ij], sizeof(float));
                if (runtimeData[netId].maxSynWt != NULL) {
                    // only copy maxSynWt if datastructure actually exists on the CPU runtime
                    // (that logic should be done elsewhere though)
                    memcpy(&runtimeData[netId].maxSynWt[pos_ij], &managerRuntimeData.maxSynWt[pos_ij], sizeof(float));
                }
            }
 
            // synapse found and updated: we're done!
            synapseFound = true;
            break;
        }
    }
 
    if (!synapseFound) {
        KERNEL_WARN("setWeight(%d,%d,%d,%f,%s): Synapse does not exist, not updated.", connId, neurIdPre, neurIdPost,
            weight, (updateWeightRange?"true":"false"));
    }
}
 
void SNN::setExternalCurrent(int grpId, const std::vector<float>& current) {
    assert(grpId >= 0); assert(grpId < numGroups);
    assert(!isPoissonGroup(grpId));
    assert(current.size() == getGroupNumNeurons(grpId));
 
    int netId = groupConfigMDMap[grpId].netId;
    int lGrpId = groupConfigMDMap[grpId].lGrpId;
 
    // // update flag for faster handling at run-time
    // if (count_if(current.begin(), current.end(), isGreaterThanZero)) {
    //  groupConfigs[0][grpId].WithCurrentInjection = true;
    // } else {
    //  groupConfigs[0][grpId].WithCurrentInjection = false;
    // }
 
    // store external current in array
    for (int lNId = groupConfigs[netId][lGrpId].lStartN, j = 0; lNId <= groupConfigs[netId][lGrpId].lEndN; lNId++, j++) {
        managerRuntimeData.extCurrent[lNId] = current[j];
    }
 
    // copy to GPU if necessary
    // don't allocate; allocation done in generateRuntimeData
    if (netId < CPU_RUNTIME_BASE) {
        copyExternalCurrent(netId, lGrpId, &runtimeData[netId], cudaMemcpyHostToDevice, false);
    }
    else {
        copyExternalCurrent(netId, lGrpId, &runtimeData[netId], false);
    }
}
 
// writes network state to file
// handling of file pointer should be handled externally: as far as this function is concerned, it is simply
// trying to write to file
void SNN::saveSimulation(FILE* fid, bool saveSynapseInfo) {
    int tmpInt;
    float tmpFloat;
 
 
    tmpInt = 294338571; // some int used to identify saveSimulation files
    if (!fwrite(&tmpInt,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
 
    tmpFloat = 0.3f;
    if (!fwrite(&tmpFloat,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
 
    tmpFloat = ((float)simTimeSec) + ((float)simTimeMs)/1000.0f;
    if (!fwrite(&tmpFloat,sizeof(float),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
 
    stopTiming();
    tmpFloat = executionTime/1000.0f;
    if (!fwrite(&tmpFloat,sizeof(float),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
 
 
    if (!fwrite(&glbNetworkConfig.numN,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    int dummyInt = 0;
    //if (!fwrite(&numPreSynNet,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    // if (!fwrite(&dummyInt,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    //if (!fwrite(&numPostSynNet,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    // if (!fwrite(&dummyInt,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    if (!fwrite(&numGroups,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
 
    char name[100];
    for (int gGrpId=0;gGrpId<numGroups;gGrpId++) {
        if (!fwrite(&groupConfigMDMap[gGrpId].gStartN,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
        if (!fwrite(&groupConfigMDMap[gGrpId].gEndN,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
 
        if (!fwrite(&groupConfigMap[gGrpId].grid.numX,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
        if (!fwrite(&groupConfigMap[gGrpId].grid.numY,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
        if (!fwrite(&groupConfigMap[gGrpId].grid.numZ,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
 
        strncpy(name,groupConfigMap[gGrpId].grpName.c_str(),100);
        if (!fwrite(name,1,100,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    }
 
    if (!saveSynapseInfo) return;
 
    // Save number of local networks
    int net_count = 0;
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            net_count++;
        }
    }
    if (!fwrite(&net_count, sizeof(int), 1, fid)) KERNEL_ERROR("saveSimulation fwrite error");
 
    // Save weights for each local network
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            // copy from runtimeData to managerRuntimeData
            fetchPreConnectionInfo(netId);
            fetchPostConnectionInfo(netId);
            fetchConnIdsLookupArray(netId);
            fetchSynapseState(netId);
 
            // save number of synapses that starting from local groups
            int numSynToSave = 0;
            for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[netId].begin(); grpIt != groupPartitionLists[netId].end(); grpIt++) {
                if (grpIt->netId == netId) {
                    numSynToSave += grpIt->numPostSynapses;
                }
            }
            if (!fwrite(&numSynToSave, sizeof(int), 1, fid)) KERNEL_ERROR("saveSimulation fwrite error");
            // read synapse info from managerRuntimData
            int numSynSaved = 0;
            for (int lNId = 0; lNId < networkConfigs[netId].numNAssigned; lNId++) {
                unsigned int offset = managerRuntimeData.cumulativePost[lNId];
 
                // save each synapse starting from from neuron lNId
                for (int t = 0; t < glbNetworkConfig.maxDelay; t++) {
                    DelayInfo dPar = managerRuntimeData.postDelayInfo[lNId*(glbNetworkConfig.maxDelay + 1)+t];
 
                    for (int idx_d=dPar.delay_index_start; idx_d < (dPar.delay_index_start + dPar.delay_length); idx_d++) {
                        SynInfo post_info = managerRuntimeData.postSynapticIds[offset + idx_d];
                        int lNIdPost = GET_CONN_NEURON_ID(post_info);
                        int lGrpIdPost = GET_CONN_GRP_ID(post_info);
                        int preSynId = GET_CONN_SYN_ID(post_info);
                        int pre_pos = managerRuntimeData.cumulativePre[lNIdPost] + preSynId;
                        SynInfo pre_info = managerRuntimeData.preSynapticIds[pre_pos];
                        int lNIdPre = GET_CONN_NEURON_ID(pre_info);
                        int lGrpIdPre = GET_CONN_GRP_ID(pre_info);
                        float weight = managerRuntimeData.wt[pre_pos];
                        float maxWeight = managerRuntimeData.maxSynWt[pre_pos];
                        int connId = managerRuntimeData.connIdsPreIdx[pre_pos];
                        int delay = t+1;
 
                        // convert local group id to global group id
                        // convert local neuron id to neuron order in group
                        int gGrpIdPre = groupConfigs[netId][lGrpIdPre].gGrpId;
                        int gGrpIdPost = groupConfigs[netId][lGrpIdPost].gGrpId;
                        int grpNIdPre = lNId - groupConfigs[netId][lGrpIdPre].lStartN;
                        int grpNIdPost = lNIdPost - groupConfigs[netId][lGrpIdPost].lStartN;
 
                        // we only save synapses starting from local groups since otherwise we will save external synapses twice 
                        // write order is based on function connectNeurons (no NetId & external_NetId)
                        // inline void SNN::connectNeurons(int netId, int _grpSrc, int _grpDest, int _nSrc, int _nDest, short int _connId, float initWt, float maxWt, uint8_t delay, int externalNetId) 
                        if (groupConfigMDMap[gGrpIdPre].netId == netId) {
                            numSynSaved++;
                            if (!fwrite(&gGrpIdPre, sizeof(int), 1, fid)) KERNEL_ERROR("saveSimulation fwrite error");
                            if (!fwrite(&gGrpIdPost, sizeof(int), 1, fid)) KERNEL_ERROR("saveSimulation fwrite error");
                            if (!fwrite(&grpNIdPre, sizeof(int), 1, fid)) KERNEL_ERROR("saveSimulation fwrite error");
                            if (!fwrite(&grpNIdPost, sizeof(int), 1, fid)) KERNEL_ERROR("saveSimulation fwrite error");
                            if (!fwrite(&connId, sizeof(int), 1, fid)) KERNEL_ERROR("saveSimulation fwrite error");
                            if (!fwrite(&weight, sizeof(float), 1, fid)) KERNEL_ERROR("saveSimulation fwrite error");
                            if (!fwrite(&maxWeight, sizeof(float), 1, fid)) KERNEL_ERROR("saveSimulation fwrite error");
                            if (!fwrite(&delay, sizeof(int), 1, fid)) KERNEL_ERROR("saveSimulation fwrite error");
                        }
                    }
                }
            }
            assert(numSynSaved == numSynToSave);
        }
    }
 
 
    //if (simMode_ == GPU_MODE)
    //  copyWeightState(&managerRuntimeData, &runtimeData[0], cudaMemcpyDeviceToHost, false);
 
    //if (saveSynapseInfo) {
    //  for (int i = 0; i < numN; i++) {
    //      unsigned int offset = managerRuntimeData.cumulativePost[i];
 
    //      unsigned int count = 0;
    //      for (int t=0;t<maxDelay_;t++) {
    //          DelayInfo dPar = managerRuntimeData.postDelayInfo[i*(maxDelay_+1)+t];
 
    //          for(int idx_d=dPar.delay_index_start; idx_d<(dPar.delay_index_start+dPar.delay_length); idx_d++)
    //              count++;
    //      }
 
    //      if (!fwrite(&count,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
 
    //      for (int t=0;t<maxDelay_;t++) {
    //          DelayInfo dPar = managerRuntimeData.postDelayInfo[i*(maxDelay_+1)+t];
 
    //          for(int idx_d=dPar.delay_index_start; idx_d<(dPar.delay_index_start+dPar.delay_length); idx_d++) {
    //              // get synaptic info...
    //              SynInfo post_info = managerRuntimeData.postSynapticIds[offset + idx_d];
 
    //              // get neuron id
    //              //int p_i = (post_info&POST_SYN_NEURON_MASK);
    //              unsigned int p_i = GET_CONN_NEURON_ID(post_info);
    //              assert(p_i<numN);
 
    //              // get syn id
    //              unsigned int s_i = GET_CONN_SYN_ID(post_info);
    //              //>>POST_SYN_NEURON_BITS)&POST_SYN_CONN_MASK;
    //              assert(s_i<(managerRuntimeData.Npre[p_i]));
 
    //              // get the cumulative position for quick access...
    //              unsigned int pos_i = managerRuntimeData.cumulativePre[p_i] + s_i;
 
    //              uint8_t delay = t+1;
    //              uint8_t plastic = s_i < managerRuntimeData.Npre_plastic[p_i]; // plastic or fixed.
 
    //              if (!fwrite(&i,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    //              if (!fwrite(&p_i,sizeof(int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    //              if (!fwrite(&(managerRuntimeData.wt[pos_i]),sizeof(float),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    //              if (!fwrite(&(managerRuntimeData.maxSynWt[pos_i]),sizeof(float),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    //              if (!fwrite(&delay,sizeof(uint8_t),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    //              if (!fwrite(&plastic,sizeof(uint8_t),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    //              if (!fwrite(&(managerRuntimeData.connIdsPreIdx[pos_i]),sizeof(short int),1,fid)) KERNEL_ERROR("saveSimulation fwrite error");
    //          }
    //      }
    //  }
    //}
}
 
// writes population weights from gIDpre to gIDpost to file fname in binary
//void SNN::writePopWeights(std::string fname, int grpIdPre, int grpIdPost) {
//  assert(grpIdPre>=0); assert(grpIdPost>=0);
//
//  float* weights;
//  int matrixSize;
//  FILE* fid;
//  int numPre, numPost;
//  fid = fopen(fname.c_str(), "wb");
//  assert(fid != NULL);
//
//  if(snnState == CONFIG_SNN || snnState == COMPILED_SNN || snnState == PARTITIONED_SNN){
//      KERNEL_ERROR("Simulation has not been run yet, cannot output weights.");
//      exitSimulation(1);
//  }
//
//  SynInfo* preId;
//  int pre_nid, pos_ij;
//
//  //population sizes
//  numPre = groupConfigs[0][grpIdPre].SizeN;
//  numPost = groupConfigs[0][grpIdPost].SizeN;
//
//  //first iteration gets the number of synaptic weights to place in our
//  //weight matrix.
//  matrixSize=0;
//  //iterate over all neurons in the post group
//  for (int i=groupConfigs[0][grpIdPost].StartN; i<=groupConfigs[0][grpIdPost].EndN; i++) {
//      // for every post-neuron, find all pre
//      pos_ij = managerRuntimeData.cumulativePre[i]; // i-th neuron, j=0th synapse
//      //iterate over all presynaptic synapses
//      for(int j=0; j<managerRuntimeData.Npre[i]; pos_ij++,j++) {
//          preId = &managerRuntimeData.preSynapticIds[pos_ij];
//          pre_nid = GET_CONN_NEURON_ID((*preId)); // neuron id of pre
//          if (pre_nid<groupConfigs[0][grpIdPre].StartN || pre_nid>groupConfigs[0][grpIdPre].EndN)
//              continue; // connection does not belong to group grpIdPre
//          matrixSize++;
//      }
//  }
//
//  //now we have the correct size
//  weights = new float[matrixSize];
//  //second iteration assigns the weights
//  int curr = 0; // iterator for return array
//  //iterate over all neurons in the post group
//  for (int i=groupConfigs[0][grpIdPost].StartN; i<=groupConfigs[0][grpIdPost].EndN; i++) {
//      // for every post-neuron, find all pre
//      pos_ij = managerRuntimeData.cumulativePre[i]; // i-th neuron, j=0th synapse
//      //do the GPU copy here.  Copy the current weights from GPU to CPU.
//      if(simMode_==GPU_MODE){
//          copyWeightsGPU(i,grpIdPre);
//      }
//      //iterate over all presynaptic synapses
//      for(int j=0; j<managerRuntimeData.Npre[i]; pos_ij++,j++) {
//          preId = &(managerRuntimeData.preSynapticIds[pos_ij]);
//          pre_nid = GET_CONN_NEURON_ID((*preId)); // neuron id of pre
//          if (pre_nid<groupConfigs[0][grpIdPre].StartN || pre_nid>groupConfigs[0][grpIdPre].EndN)
//              continue; // connection does not belong to group grpIdPre
//          weights[curr] = managerRuntimeData.wt[pos_ij];
//          curr++;
//      }
//  }
//
//  fwrite(weights,sizeof(float),matrixSize,fid);
//  fclose(fid);
//  //Let my memory FREE!!!
//  delete [] weights;
//}
 
 
 
// set new file pointer for all files
// fp==NULL is code for don't change it
// can be called in all logger modes; however, the analogous interface function can only be called in CUSTOM
void SNN::setLogsFp(FILE* fpInf, FILE* fpErr, FILE* fpDeb, FILE* fpLog) {
    if (fpInf!=NULL) {
        if (fpInf_!=NULL && fpInf_!=stdout && fpInf_!=stderr)
            fclose(fpInf_);
        fpInf_ = fpInf;
    }
 
    if (fpErr!=NULL) {
        if (fpErr_ != NULL && fpErr_!=stdout && fpErr_!=stderr)
            fclose(fpErr_);
        fpErr_ = fpErr;
    }
 
    if (fpDeb!=NULL) {
        if (fpDeb_!=NULL && fpDeb_!=stdout && fpDeb_!=stderr)
            fclose(fpDeb_);
        fpDeb_ = fpDeb;
    }
 
    if (fpLog!=NULL) {
        if (fpLog_!=NULL && fpLog_!=stdout && fpLog_!=stderr)
            fclose(fpLog_);
        fpLog_ = fpLog;
    }
}
 
 
 
// loop over linked list entries to find a connection with the right pre-post pair, O(N)
short int SNN::getConnectId(int grpIdPre, int grpIdPost) {
    short int connId = -1;
 
    for (std::map<int, ConnectConfig>::iterator it = connectConfigMap.begin(); it != connectConfigMap.end(); it++) {
        if (it->second.grpSrc == grpIdPre && it->second.grpDest == grpIdPost) {
            connId = it->second.connId;
            break;
        }
    }
 
    return connId;
}
 
ConnectConfig SNN::getConnectConfig(short int connId) {
    CHECK_CONNECTION_ID(connId, numConnections);
 
    if (connectConfigMap.find(connId) == connectConfigMap.end()) {
        KERNEL_ERROR("Total Connections = %d", numConnections);
        KERNEL_ERROR("ConnectId (%d) cannot be recognized", connId);
    }
 
    return connectConfigMap[connId];
}
 
std::vector<float> SNN::getConductanceAMPA(int gGrpId) {
    assert(isSimulationWithCOBA());
 
    // copy data to the manager runtime
    fetchConductanceAMPA(gGrpId);
 
    std::vector<float> gAMPAvec;
    for (int gNId = groupConfigMDMap[gGrpId].gStartN; gNId <= groupConfigMDMap[gGrpId].gEndN; gNId++) {
        gAMPAvec.push_back(managerRuntimeData.gAMPA[gNId]);
    }
    return gAMPAvec;
}
 
std::vector<float> SNN::getConductanceNMDA(int gGrpId) {
    assert(isSimulationWithCOBA());
 
    // copy data to the manager runtime
    fetchConductanceNMDA(gGrpId);
 
    std::vector<float> gNMDAvec;
    if (isSimulationWithNMDARise()) {
        // need to construct conductance from rise and decay parts
        for (int gNId = groupConfigMDMap[gGrpId].gStartN; gNId <= groupConfigMDMap[gGrpId].gEndN; gNId++) {
            gNMDAvec.push_back(managerRuntimeData.gNMDA_d[gNId] - managerRuntimeData.gNMDA_r[gNId]);
        }
    } else {
        for (int gNId = groupConfigMDMap[gGrpId].gStartN; gNId <= groupConfigMDMap[gGrpId].gEndN; gNId++) {
            gNMDAvec.push_back(managerRuntimeData.gNMDA[gNId]);
        }
    }
    return gNMDAvec;
}
 
std::vector<float> SNN::getConductanceGABAa(int gGrpId) {
    assert(isSimulationWithCOBA());
 
    // copy data to the manager runtime
    fetchConductanceGABAa(gGrpId);
 
    std::vector<float> gGABAaVec;
    for (int gNId = groupConfigMDMap[gGrpId].gStartN; gNId <= groupConfigMDMap[gGrpId].gEndN; gNId++) {
        gGABAaVec.push_back(managerRuntimeData.gGABAa[gNId]);
    }
    return gGABAaVec;
}
 
std::vector<float> SNN::getConductanceGABAb(int gGrpId) {
    assert(isSimulationWithCOBA());
 
    // copy data to the manager runtime
    fetchConductanceGABAb(gGrpId);
 
    std::vector<float> gGABAbVec;
    if (isSimulationWithGABAbRise()) {
        // need to construct conductance from rise and decay parts
        for (int gNId = groupConfigMDMap[gGrpId].gStartN; gNId <= groupConfigMDMap[gGrpId].gEndN; gNId++) {
            gGABAbVec.push_back(managerRuntimeData.gGABAb_d[gNId] - managerRuntimeData.gGABAb_r[gNId]);
        }
    } else {
        for (int gNId = groupConfigMDMap[gGrpId].gStartN; gNId <= groupConfigMDMap[gGrpId].gEndN; gNId++) {
            gGABAbVec.push_back(managerRuntimeData.gGABAb[gNId]);
        }
    }
    return gGABAbVec;
}
 
// returns RangeDelay struct of a connection
RangeDelay SNN::getDelayRange(short int connId) {
    assert(connId>=0 && connId<numConnections);
 
    return RangeDelay(connectConfigMap[connId].minDelay, connectConfigMap[connId].maxDelay);
}
 
// \TODO: bad API design (return allocated memory to user), consider to move this function to connection monitor
uint8_t* SNN::getDelays(int gGrpIdPre, int gGrpIdPost, int& numPreN, int& numPostN) {
    int netIdPost = groupConfigMDMap[gGrpIdPost].netId;
    int lGrpIdPost = groupConfigMDMap[gGrpIdPost].lGrpId;
    int lGrpIdPre = -1;
    uint8_t* delays;
 
    for (int lGrpId = 0; lGrpId < networkConfigs[netIdPost].numGroupsAssigned; lGrpId++)
        if (groupConfigs[netIdPost][lGrpId].gGrpId == gGrpIdPre) {
            lGrpIdPre = lGrpId;
            break;
        }
    assert(lGrpIdPre != -1);
 
    numPreN = groupConfigMap[gGrpIdPre].numN;
    numPostN = groupConfigMap[gGrpIdPost].numN;
 
    delays = new uint8_t[numPreN * numPostN];
    memset(delays, 0, numPreN * numPostN);
 
    fetchPostConnectionInfo(netIdPost);
 
    for (int lNIdPre = groupConfigs[netIdPost][lGrpIdPre].lStartN; lNIdPre < groupConfigs[netIdPost][lGrpIdPre].lEndN; lNIdPre++) {
        unsigned int offset = managerRuntimeData.cumulativePost[lNIdPre];
 
        for (int t = 0; t < glbNetworkConfig.maxDelay; t++) {
            DelayInfo dPar = managerRuntimeData.postDelayInfo[lNIdPre * (glbNetworkConfig.maxDelay + 1) + t];
 
            for(int idx_d = dPar.delay_index_start; idx_d<(dPar.delay_index_start+dPar.delay_length); idx_d++) {
                // get synaptic info...
                SynInfo postSynInfo = managerRuntimeData.postSynapticIds[offset + idx_d];
 
                // get local post neuron id
                int lNIdPost = GET_CONN_NEURON_ID(postSynInfo);
                assert(lNIdPost < glbNetworkConfig.numN);
 
                if (lNIdPost >= groupConfigs[netIdPost][lGrpIdPost].lStartN && lNIdPost <= groupConfigs[netIdPost][lGrpIdPost].lEndN) {
                    delays[(lNIdPre - groupConfigs[netIdPost][lGrpIdPre].lStartN) + numPreN * (lNIdPost - groupConfigs[netIdPost][lGrpIdPost].lStartN)] = t + 1;
                }
            }
        }
    }
    return delays;
}
 
Grid3D SNN::getGroupGrid3D(int gGrpId) {
    assert(gGrpId >= 0 && gGrpId < numGroups);
 
    return groupConfigMap[gGrpId].grid;
}
 
// find ID of group with name grpName
int SNN::getGroupId(std::string grpName) {
    int grpId = -1;
    for (int gGrpId = 0; gGrpId < numGroups; gGrpId++) {
        if (groupConfigMap[gGrpId].grpName.compare(grpName) == 0) {
            grpId = gGrpId;
            break;
        }
    }
 
    return grpId;
}
 
std::string SNN::getGroupName(int gGrpId) {
    assert(gGrpId >= -1 && gGrpId < numGroups);
 
    if (gGrpId == ALL)
        return "ALL";
 
    return groupConfigMap[gGrpId].grpName;
}
 
GroupSTDPInfo SNN::getGroupSTDPInfo(int gGrpId) {
    GroupSTDPInfo gInfo;
 
    gInfo.WithSTDP = groupConfigMap[gGrpId].stdpConfig.WithSTDP;
    gInfo.WithESTDP = groupConfigMap[gGrpId].stdpConfig.WithESTDP;
    gInfo.WithISTDP = groupConfigMap[gGrpId].stdpConfig.WithISTDP;
    gInfo.WithESTDPtype = groupConfigMap[gGrpId].stdpConfig.WithESTDPtype;
    gInfo.WithISTDPtype = groupConfigMap[gGrpId].stdpConfig.WithISTDPtype;
    gInfo.WithESTDPcurve = groupConfigMap[gGrpId].stdpConfig.WithESTDPcurve;
    gInfo.WithISTDPcurve = groupConfigMap[gGrpId].stdpConfig.WithISTDPcurve;
    gInfo.ALPHA_MINUS_EXC = groupConfigMap[gGrpId].stdpConfig.ALPHA_MINUS_EXC;
    gInfo.ALPHA_PLUS_EXC = groupConfigMap[gGrpId].stdpConfig.ALPHA_PLUS_EXC;
    gInfo.TAU_MINUS_INV_EXC = groupConfigMap[gGrpId].stdpConfig.TAU_MINUS_INV_EXC;
    gInfo.TAU_PLUS_INV_EXC = groupConfigMap[gGrpId].stdpConfig.TAU_PLUS_INV_EXC;
    gInfo.ALPHA_MINUS_INB = groupConfigMap[gGrpId].stdpConfig.ALPHA_MINUS_INB;
    gInfo.ALPHA_PLUS_INB = groupConfigMap[gGrpId].stdpConfig.ALPHA_PLUS_INB;
    gInfo.TAU_MINUS_INV_INB = groupConfigMap[gGrpId].stdpConfig.TAU_MINUS_INV_INB;
    gInfo.TAU_PLUS_INV_INB = groupConfigMap[gGrpId].stdpConfig.TAU_PLUS_INV_INB;
    gInfo.GAMMA = groupConfigMap[gGrpId].stdpConfig.GAMMA;
    gInfo.BETA_LTP = groupConfigMap[gGrpId].stdpConfig.BETA_LTP;
    gInfo.BETA_LTD = groupConfigMap[gGrpId].stdpConfig.BETA_LTD;
    gInfo.LAMBDA = groupConfigMap[gGrpId].stdpConfig.LAMBDA;
    gInfo.DELTA = groupConfigMap[gGrpId].stdpConfig.DELTA;
 
    return gInfo;
}
 
GroupNeuromodulatorInfo SNN::getGroupNeuromodulatorInfo(int gGrpId) {
    GroupNeuromodulatorInfo gInfo;
 
    gInfo.baseDP = groupConfigMap[gGrpId].neuromodulatorConfig.baseDP;
    gInfo.base5HT = groupConfigMap[gGrpId].neuromodulatorConfig.base5HT;
    gInfo.baseACh = groupConfigMap[gGrpId].neuromodulatorConfig.baseACh;
    gInfo.baseNE = groupConfigMap[gGrpId].neuromodulatorConfig.baseNE;
    gInfo.decayDP = groupConfigMap[gGrpId].neuromodulatorConfig.decayDP;
    gInfo.decay5HT = groupConfigMap[gGrpId].neuromodulatorConfig.decay5HT;
    gInfo.decayACh = groupConfigMap[gGrpId].neuromodulatorConfig.decayACh;
    gInfo.decayNE = groupConfigMap[gGrpId].neuromodulatorConfig.decayNE;
 
    return gInfo;
}
 
Point3D SNN::getNeuronLocation3D(int gNId) {
    int gGrpId = -1;
    assert(gNId >= 0 && gNId < glbNetworkConfig.numN);
 
    // search for global group id
    for (std::map<int, GroupConfigMD>::iterator grpIt = groupConfigMDMap.begin(); grpIt != groupConfigMDMap.end(); grpIt++) {
        if (gNId >= grpIt->second.gStartN && gNId <= grpIt->second.gEndN)
            gGrpId = grpIt->second.gGrpId;
    }
 
    // adjust neurId for neuron ID of first neuron in the group
    int neurId = gNId - groupConfigMDMap[gGrpId].gStartN;
 
    return getNeuronLocation3D(gGrpId, neurId);
}
 
Point3D SNN::getNeuronLocation3D(int gGrpId, int relNeurId) {
    Grid3D grid = groupConfigMap[gGrpId].grid;
    assert(gGrpId >= 0 && gGrpId < numGroups);
    assert(relNeurId >= 0 && relNeurId < getGroupNumNeurons(gGrpId));
 
    int intX = relNeurId % grid.numX;
    int intY = (relNeurId / grid.numX) % grid.numY;
    int intZ = relNeurId / (grid.numX * grid.numY);
 
    // get coordinates center around origin
    double coordX = grid.distX * intX + grid.offsetX;
    double coordY = grid.distY * intY + grid.offsetY;
    double coordZ = grid.distZ * intZ + grid.offsetZ;
    return Point3D(coordX, coordY, coordZ);
}
 
// returns the number of synaptic connections associated with this connection.
int SNN::getNumSynapticConnections(short int connId) {
    //we didn't find the connection.
    if (connectConfigMap.find(connId) == connectConfigMap.end()) {
        KERNEL_ERROR("Connection ID was not found.  Quitting.");
        exitSimulation(1);
    }
 
    return connectConfigMap[connId].numberOfConnections;
}
 
// returns pointer to existing SpikeMonitor object, NULL else
SpikeMonitor* SNN::getSpikeMonitor(int gGrpId) {
    assert(gGrpId >= 0 && gGrpId < getNumGroups());
 
    if (groupConfigMDMap[gGrpId].spikeMonitorId >= 0) {
        return spikeMonList[(groupConfigMDMap[gGrpId].spikeMonitorId)];
    } else {
        return NULL;
    }
}
 
SpikeMonitorCore* SNN::getSpikeMonitorCore(int gGrpId) {
    assert(gGrpId >= 0 && gGrpId < getNumGroups());
 
    if (groupConfigMDMap[gGrpId].spikeMonitorId >= 0) {
        return spikeMonCoreList[(groupConfigMDMap[gGrpId].spikeMonitorId)];
    } else {
        return NULL;
    }
}
 
// returns pointer to existing NeuronMonitor object, NULL else
NeuronMonitor* SNN::getNeuronMonitor(int gGrpId) {
    assert(gGrpId >= 0 && gGrpId < getNumGroups());
 
    if (groupConfigMDMap[gGrpId].neuronMonitorId >= 0) {
        return neuronMonList[(groupConfigMDMap[gGrpId].neuronMonitorId)];
    }
    else {
        return NULL;
    }
}
 
NeuronMonitorCore* SNN::getNeuronMonitorCore(int gGrpId) {
    assert(gGrpId >= 0 && gGrpId < getNumGroups());
 
    if (groupConfigMDMap[gGrpId].neuronMonitorId >= 0) {
        return neuronMonCoreList[(groupConfigMDMap[gGrpId].neuronMonitorId)];
    }
    else {
        return NULL;
    }
}
 
RangeWeight SNN::getWeightRange(short int connId) {
    assert(connId>=0 && connId<numConnections);
 
    return RangeWeight(0.0f, connectConfigMap[connId].initWt, connectConfigMap[connId].maxWt);
}
 
 
 
// all unsafe operations of SNN constructor
void SNN::SNNinit() {
    // initialize snnState
    snnState = CONFIG_SNN;
 
    // set logger mode (defines where to print all status, error, and debug messages)
    switch (loggerMode_) {
    case USER:
        fpInf_ = stdout;
        fpErr_ = stderr;
        #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
            fpDeb_ = fopen("nul","w");
        #else
            fpDeb_ = fopen("/dev/null","w");
        #endif
        break;
    case DEVELOPER:
        fpInf_ = stdout;
        fpErr_ = stderr;
        fpDeb_ = stdout;
        break;
    case SHOWTIME:
        #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
            fpInf_ = fopen("nul","w");
        #else
            fpInf_ = fopen("/dev/null","w");
        #endif
        fpErr_ = stderr;
        #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
            fpDeb_ = fopen("nul","w");
        #else
            fpDeb_ = fopen("/dev/null","w");
        #endif
        break;
    case SILENT:
    case CUSTOM:
        #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
            fpInf_ = fopen("nul","w");
            fpErr_ = fopen("nul","w");
            fpDeb_ = fopen("nul","w");
        #else
            fpInf_ = fopen("/dev/null","w");
            fpErr_ = fopen("/dev/null","w");
            fpDeb_ = fopen("/dev/null","w");
        #endif
    break;
    default:
        fpErr_ = stderr; // need to open file stream first
        KERNEL_ERROR("Unknown logger mode");
        exit(UNKNOWN_LOGGER_ERROR);
 
    }
 
    // try to open log file in results folder: create if not exists
#if defined(WIN32) || defined(WIN64)
    CreateDirectory("results", NULL);
    fpLog_ = fopen("results/carlsim.log", "w");
#else
    struct stat sb;
    int createDir = 1;
    if (stat("results", &sb) == -1 || !S_ISDIR(sb.st_mode)) {
        // results dir does not exist, try to create:
        createDir = mkdir("results", 0777);
    }
 
    if (createDir == -1) {
        // tried to create dir, but failed
        fprintf(stderr, "Could not create directory \"results/\", which is required to "
            "store simulation results. Aborting simulation...\n");
        exit(NO_LOGGER_DIR_ERROR);
    } else {
        // open log file
        fpLog_ = fopen("results/carlsim.log", "w");
 
        if (createDir == 0) {
            // newly created dir: now that fpLog_/fpInf_ exist, inform user
            KERNEL_INFO("Created results directory \"results/\".");
        }
    }
#endif
    if (fpLog_ == NULL) {
        fprintf(stderr, "Could not create the directory \"results/\" or the log file \"results/carlsim.log\""
            ", which is required to store simulation results. Aborting simulation...\n");
        exit(NO_LOGGER_DIR_ERROR);
    }
 
    KERNEL_INFO("*********************************************************************************");
    KERNEL_INFO("********************      Welcome to CARLsim %d.%d      ***************************",
                MAJOR_VERSION,MINOR_VERSION);
    KERNEL_INFO("*********************************************************************************\n");
 
    KERNEL_INFO("***************************** Configuring Network ********************************");
    KERNEL_INFO("Starting CARLsim simulation \"%s\" in %s mode",networkName_.c_str(),
        loggerMode_string[loggerMode_]);
    KERNEL_INFO("Random number seed: %d",randSeed_);
 
    time_t rawtime;
    struct tm * timeinfo;
    time(&rawtime);
    timeinfo = localtime(&rawtime);
    KERNEL_DEBUG("Current local time and date: %s", asctime(timeinfo));
 
    // init random seed
    srand48(randSeed_);
 
    simTimeRunStart = 0; simTimeRunStop = 0;
    simTimeLastRunSummary = 0;
    simTimeMs = 0; simTimeSec = 0; simTime = 0;
 
    numGroups = 0;
    numConnections = 0;
    numCompartmentConnections = 0;
    numSpikeGenGrps = 0;
    simulatorDeleted = false;
 
    cumExecutionTime = 0.0;
    executionTime = 0.0;
 
    spikeRateUpdated = false;
    numSpikeMonitor = 0;
    numNeuronMonitor = 0;
    numGroupMonitor = 0;
    numConnectionMonitor = 0;
 
    sim_with_compartments = false;
    sim_with_fixedwts = true; // default is true, will be set to false if there are any plastic synapses
    sim_with_conductances = false; // default is false
    sim_with_stdp = false;
    sim_with_modulated_stdp = false;
    sim_with_homeostasis = false;
    sim_with_stp = false;
    sim_in_testing = false;
 
    loadSimFID = NULL;
 
    // conductance info struct for simulation
    sim_with_NMDA_rise = false;
    sim_with_GABAb_rise = false;
    dAMPA  = 1.0-1.0/5.0;       // some default decay and rise times
    rNMDA  = 1.0-1.0/10.0;
    dNMDA  = 1.0-1.0/150.0;
    sNMDA  = 1.0;
    dGABAa = 1.0-1.0/6.0;
    rGABAb = 1.0-1.0/100.0;
    dGABAb = 1.0-1.0/150.0;
    sGABAb = 1.0;
 
    // default integration method: Forward-Euler with 0.5ms integration step
    setIntegrationMethod(FORWARD_EULER, 2);
 
    mulSynFast = NULL;
    mulSynSlow = NULL;
 
    // reset all monitors, don't deallocate (false)
    resetMonitors(false);
 
    resetGroupConfigs(false);
 
    resetConnectionConfigs(false);
 
    // initialize spike buffer
    spikeBuf = new SpikeBuffer(0, MAX_TIME_SLICE);
 
    memset(networkConfigs, 0, sizeof(NetworkConfigRT) * MAX_NET_PER_SNN);
 
    // reset all runtime data
    // GPU/CPU runtime data
    memset(runtimeData, 0, sizeof(RuntimeData) * MAX_NET_PER_SNN);
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) // FIXME: redundant??
        runtimeData[netId].allocated = false;
 
    // Manager runtime data
    memset(&managerRuntimeData, 0, sizeof(RuntimeData));
    managerRuntimeData.allocated = false; // FIXME: redundant??
 
    // default weight update parameter
    wtANDwtChangeUpdateInterval_ = 1000; // update weights every 1000 ms (default)
    wtANDwtChangeUpdateIntervalCnt_ = 0; // helper var to implement fast modulo
    stdpScaleFactor_ = 1.0f;
    wtChangeDecay_ = 0.0f;
 
    // FIXME: use it when necessary
#ifndef __NO_CUDA__
    CUDA_CREATE_TIMER(timer);
    CUDA_RESET_TIMER(timer);
#endif
}
 
void SNN::advSimStep() {
    doSTPUpdateAndDecayCond();
 
    //KERNEL_INFO("STPUpdate!");
 
    spikeGeneratorUpdate();
 
    //KERNEL_INFO("spikeGeneratorUpdate!");
 
    findFiring();
 
    //KERNEL_INFO("Find firing!");
 
    updateTimingTable();
 
    routeSpikes();
 
    doCurrentUpdate();
 
    //KERNEL_INFO("doCurrentUpdate!");
 
    globalStateUpdate();
 
    //KERNEL_INFO("globalStateUpdate!");
 
    clearExtFiringTable();
}
 
void SNN::doSTPUpdateAndDecayCond() {
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
        cpu_set_t cpus;
        ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
        int threadCount = 0;
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            assert(runtimeData[netId].allocated);
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                doSTPUpdateAndDecayCond_GPU(netId);
            else{//CPU runtime
                #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                    doSTPUpdateAndDecayCond_CPU(netId);
                #else // Linux or MAC
                    pthread_attr_t attr;
                    pthread_attr_init(&attr);
                    CPU_ZERO(&cpus);
                    CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                    pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                    argsThreadRoutine[threadCount].snn_pointer = this;
                    argsThreadRoutine[threadCount].netId = netId;
                    argsThreadRoutine[threadCount].lGrpId = 0;
                    argsThreadRoutine[threadCount].startIdx = 0;
                    argsThreadRoutine[threadCount].endIdx = 0;
                    argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                    pthread_create(&threads[threadCount], &attr, &SNN::helperDoSTPUpdateAndDecayCond_CPU, (void*)&argsThreadRoutine[threadCount]);
                    pthread_attr_destroy(&attr);
                    threadCount++;
                #endif
            }
        }
    }
 
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        // join all the threads
        for (int i=0; i<threadCount; i++){
            pthread_join(threads[i], NULL);
        }
    #endif
}
 
void SNN::spikeGeneratorUpdate() {
    // If poisson rate has been updated, assign new poisson rate
    if (spikeRateUpdated) {
        #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
            pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
            cpu_set_t cpus;
            ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
            int threadCount = 0;
        #endif
 
        for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
            if (!groupPartitionLists[netId].empty()) {
                if (netId < CPU_RUNTIME_BASE) // GPU runtime
                    assignPoissonFiringRate_GPU(netId);
                else{ // CPU runtime
                    #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                        assignPoissonFiringRate_CPU(netId);
                    #else // Linux or MAC
                        pthread_attr_t attr;
                        pthread_attr_init(&attr);
                        CPU_ZERO(&cpus);
                        CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                        pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                        argsThreadRoutine[threadCount].snn_pointer = this;
                        argsThreadRoutine[threadCount].netId = netId;
                        argsThreadRoutine[threadCount].lGrpId = 0;
                        argsThreadRoutine[threadCount].startIdx = 0;
                        argsThreadRoutine[threadCount].endIdx = 0;
                        argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                        pthread_create(&threads[threadCount], &attr, &SNN::helperAssignPoissonFiringRate_CPU, (void*)&argsThreadRoutine[threadCount]);
                        pthread_attr_destroy(&attr);
                        threadCount++;
                    #endif
                }
            }
        }
 
        #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
            // join all the threads
            for (int i=0; i<threadCount; i++){
                pthread_join(threads[i], NULL);
            }
        #endif
 
        spikeRateUpdated = false;
    }
 
    // If time slice has expired, check if new spikes needs to be generated by user-defined spike generators
    generateUserDefinedSpikes();
 
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
        cpu_set_t cpus;
        ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
        int threadCount = 0;
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                spikeGeneratorUpdate_GPU(netId);
            else{ // CPU runtime
                #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                    spikeGeneratorUpdate_CPU(netId);
                #else // Linux or MAC
                    pthread_attr_t attr;
                    pthread_attr_init(&attr);
                    CPU_ZERO(&cpus);
                    CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                    pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                    argsThreadRoutine[threadCount].snn_pointer = this;
                    argsThreadRoutine[threadCount].netId = netId;
                    argsThreadRoutine[threadCount].lGrpId = 0;
                    argsThreadRoutine[threadCount].startIdx = 0;
                    argsThreadRoutine[threadCount].endIdx = 0;
                    argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                    pthread_create(&threads[threadCount], &attr, &SNN::helperSpikeGeneratorUpdate_CPU, (void*)&argsThreadRoutine[threadCount]);
                    pthread_attr_destroy(&attr);
                    threadCount++;
                #endif
            }
        }
    }
 
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        // join all the threads
        for (int i=0; i<threadCount; i++){
            pthread_join(threads[i], NULL);
        }
    #endif
 
    // tell the spike buffer to advance to the next time step
    spikeBuf->step();
}
 
void SNN::findFiring() {
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
        cpu_set_t cpus;
        ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
        int threadCount = 0;
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                findFiring_GPU(netId);
            else {// CPU runtime
                #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                    findFiring_CPU(netId);
                #else // Linux or MAC
                    pthread_attr_t attr;
                    pthread_attr_init(&attr);
                    CPU_ZERO(&cpus);
                    CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                    pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                    argsThreadRoutine[threadCount].snn_pointer = this;
                    argsThreadRoutine[threadCount].netId = netId;
                    argsThreadRoutine[threadCount].lGrpId = 0;
                    argsThreadRoutine[threadCount].startIdx = 0;
                    argsThreadRoutine[threadCount].endIdx = 0;
                    argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                    pthread_create(&threads[threadCount], &attr, &SNN::helperFindFiring_CPU, (void*)&argsThreadRoutine[threadCount]);
                    pthread_attr_destroy(&attr);
                    threadCount++;
                #endif
            }
        }
    }
 
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        // join all the threads
        for (int i=0; i<threadCount; i++){
            pthread_join(threads[i], NULL);
        }
    #endif
}
 
void SNN::doCurrentUpdate() {
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
        cpu_set_t cpus;
        ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
        int threadCount = 0;
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                doCurrentUpdateD2_GPU(netId);
            else{ // CPU runtime
                #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                    doCurrentUpdateD2_CPU(netId);
                #else // Linux or MAC
                    pthread_attr_t attr;
                    pthread_attr_init(&attr);
                    CPU_ZERO(&cpus);
                    CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                    pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                    argsThreadRoutine[threadCount].snn_pointer = this;
                    argsThreadRoutine[threadCount].netId = netId;
                    argsThreadRoutine[threadCount].lGrpId = 0;
                    argsThreadRoutine[threadCount].startIdx = 0;
                    argsThreadRoutine[threadCount].endIdx = 0;
                    argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                    pthread_create(&threads[threadCount], &attr, &SNN::helperDoCurrentUpdateD2_CPU, (void*)&argsThreadRoutine[threadCount]);
                    pthread_attr_destroy(&attr);
                    threadCount++;
                #endif
            }
        }
    }
 
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        // join all the threads
        for (int i=0; i<threadCount; i++){
            pthread_join(threads[i], NULL);
        }
        threadCount = 0;
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                doCurrentUpdateD1_GPU(netId);
            else{ // CPU runtime
                #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                    doCurrentUpdateD1_CPU(netId);
                #else // Linux or MAC
                    pthread_attr_t attr;
                    pthread_attr_init(&attr);
                    CPU_ZERO(&cpus);
                    CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                    pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                    argsThreadRoutine[threadCount].snn_pointer = this;
                    argsThreadRoutine[threadCount].netId = netId;
                    argsThreadRoutine[threadCount].lGrpId = 0;
                    argsThreadRoutine[threadCount].startIdx = 0;
                    argsThreadRoutine[threadCount].endIdx = 0;
                    argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                    pthread_create(&threads[threadCount], &attr, &SNN::helperDoCurrentUpdateD1_CPU, (void*)&argsThreadRoutine[threadCount]);
                    pthread_attr_destroy(&attr);
                    threadCount++;
                #endif
            }
        }
    }
 
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        // join all the threads
        for (int i=0; i<threadCount; i++){
            pthread_join(threads[i], NULL);
        }
    #endif
}
 
void SNN::updateTimingTable() {
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
        cpu_set_t cpus;
        ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
        int threadCount = 0;
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                updateTimingTable_GPU(netId);
            else{ // CPU runtime
                #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                    updateTimingTable_CPU(netId);
                #else // Linux or MAC
                    pthread_attr_t attr;
                    pthread_attr_init(&attr);
                    CPU_ZERO(&cpus);
                    CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                    pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                    argsThreadRoutine[threadCount].snn_pointer = this;
                    argsThreadRoutine[threadCount].netId = netId;
                    argsThreadRoutine[threadCount].lGrpId = 0;
                    argsThreadRoutine[threadCount].startIdx = 0;
                    argsThreadRoutine[threadCount].endIdx = 0;
                    argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                    pthread_create(&threads[threadCount], &attr, &SNN::helperUpdateTimingTable_CPU, (void*)&argsThreadRoutine[threadCount]);
                    pthread_attr_destroy(&attr);
                    threadCount++;
                #endif
            }
        }
    }
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        // join all the threads
        for (int i=0; i<threadCount; i++){
            pthread_join(threads[i], NULL);
        }
    #endif
}
 
void SNN::globalStateUpdate() {
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
        cpu_set_t cpus;
        ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
        int threadCount = 0;
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                globalStateUpdate_C_GPU(netId);
            else{ // CPU runtime
                #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                    globalStateUpdate_CPU(netId);
                #else // Linux or MAC
                    pthread_attr_t attr;
                    pthread_attr_init(&attr);
                    CPU_ZERO(&cpus);
                    CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                    pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                    argsThreadRoutine[threadCount].snn_pointer = this;
                    argsThreadRoutine[threadCount].netId = netId;
                    argsThreadRoutine[threadCount].lGrpId = 0;
                    argsThreadRoutine[threadCount].startIdx = 0;
                    argsThreadRoutine[threadCount].endIdx = 0;
                    argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                    pthread_create(&threads[threadCount], &attr, &SNN::helperGlobalStateUpdate_CPU, (void*)&argsThreadRoutine[threadCount]);
                    pthread_attr_destroy(&attr);
                    threadCount++;
                #endif
            }
        }
    }
 
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        // join all the threads
        for (int i=0; i<threadCount; i++){
            pthread_join(threads[i], NULL);
        }
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                globalStateUpdate_N_GPU(netId);
        }
    }
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                globalStateUpdate_G_GPU(netId);
        }
    }
}
 
void SNN::clearExtFiringTable() {
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
        cpu_set_t cpus;
        ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
        int threadCount = 0;
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                clearExtFiringTable_GPU(netId);
            else{ // CPU runtime
                #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                    clearExtFiringTable_CPU(netId);
                #else // Linux or MAC
                    pthread_attr_t attr;
                    pthread_attr_init(&attr);
                    CPU_ZERO(&cpus);
                    CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                    pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                    argsThreadRoutine[threadCount].snn_pointer = this;
                    argsThreadRoutine[threadCount].netId = netId;
                    argsThreadRoutine[threadCount].lGrpId = 0;
                    argsThreadRoutine[threadCount].startIdx = 0;
                    argsThreadRoutine[threadCount].endIdx = 0;
                    argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                    pthread_create(&threads[threadCount], &attr, &SNN::helperClearExtFiringTable_CPU, (void*)&argsThreadRoutine[threadCount]);
                    pthread_attr_destroy(&attr);
                    threadCount++;
                #endif
            }
        }
    }
 
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        // join all the threads
        for (int i=0; i<threadCount; i++){
            pthread_join(threads[i], NULL);
        }
    #endif
}
 
void SNN::updateWeights() {
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
        cpu_set_t cpus;
        ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
        int threadCount = 0;
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                updateWeights_GPU(netId);
            else{ // CPU runtime
                #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                    updateWeights_CPU(netId);
                #else // Linux or MAC
                    pthread_attr_t attr;
                    pthread_attr_init(&attr);
                    CPU_ZERO(&cpus);
                    CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                    pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                    argsThreadRoutine[threadCount].snn_pointer = this;
                    argsThreadRoutine[threadCount].netId = netId;
                    argsThreadRoutine[threadCount].lGrpId = 0;
                    argsThreadRoutine[threadCount].startIdx = 0;
                    argsThreadRoutine[threadCount].endIdx = 0;
                    argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                    pthread_create(&threads[threadCount], &attr, &SNN::helperUpdateWeights_CPU, (void*)&argsThreadRoutine[threadCount]);
                    pthread_attr_destroy(&attr);
                    threadCount++;
                #endif
            }
        }
    }
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        // join all the threads
        for (int i=0; i<threadCount; i++){
            pthread_join(threads[i], NULL);
        }
    #endif
 
}
 
void SNN::updateNetworkConfig(int netId) {
    assert(netId < MAX_NET_PER_SNN);
 
    if (netId < CPU_RUNTIME_BASE) // GPU runtime
        copyNetworkConfig(netId, cudaMemcpyHostToDevice);
    else
        copyNetworkConfig(netId); // CPU runtime
}
 
void SNN::shiftSpikeTables() {
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
        cpu_set_t cpus;
        ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
        int threadCount = 0;
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                shiftSpikeTables_F_GPU(netId);
            else { // CPU runtime
                #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                    shiftSpikeTables_CPU(netId);
                #else // Linux or MAC
                    pthread_attr_t attr;
                    pthread_attr_init(&attr);
                    CPU_ZERO(&cpus);
                    CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                    pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                    argsThreadRoutine[threadCount].snn_pointer = this;
                    argsThreadRoutine[threadCount].netId = netId;
                    argsThreadRoutine[threadCount].lGrpId = 0;
                    argsThreadRoutine[threadCount].startIdx = 0;
                    argsThreadRoutine[threadCount].endIdx = 0;
                    argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                    pthread_create(&threads[threadCount], &attr, &SNN::helperShiftSpikeTables_CPU, (void*)&argsThreadRoutine[threadCount]);
                    pthread_attr_destroy(&attr);
                    threadCount++;
                #endif
            }
        }
    }
 
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        // join all the threads
        for (int i=0; i<threadCount; i++){
            pthread_join(threads[i], NULL);
        }
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                shiftSpikeTables_T_GPU(netId);
        }
    }
}
 
void SNN::allocateSNN(int netId) {
    assert(netId > ANY && netId < MAX_NET_PER_SNN);
 
    if (netId < CPU_RUNTIME_BASE)
        allocateSNN_GPU(netId);
    else
        allocateSNN_CPU(netId);
}
 
void SNN::allocateManagerRuntimeData() {
    // reset variable related to spike count
    managerRuntimeData.spikeCountSec = 0;
    managerRuntimeData.spikeCountD1Sec = 0;
    managerRuntimeData.spikeCountD2Sec = 0;
    managerRuntimeData.spikeCountLastSecLeftD2 = 0;
    managerRuntimeData.spikeCount = 0;
    managerRuntimeData.spikeCountD1 = 0;
    managerRuntimeData.spikeCountD2 = 0;
    managerRuntimeData.nPoissonSpikes = 0;
    managerRuntimeData.spikeCountExtRxD1 = 0;
    managerRuntimeData.spikeCountExtRxD2 = 0;
 
    managerRuntimeData.voltage    = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.nextVoltage = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.recovery   = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.Izh_a      = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.Izh_b      = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.Izh_c      = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.Izh_d      = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.Izh_C      = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.Izh_k      = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.Izh_vr     = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.Izh_vt     = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.Izh_vpeak  = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.lif_tau_m      = new int[managerRTDSize.maxNumNReg];
    managerRuntimeData.lif_tau_ref      = new int[managerRTDSize.maxNumNReg];
    managerRuntimeData.lif_tau_ref_c      = new int[managerRTDSize.maxNumNReg];
    managerRuntimeData.lif_vTh      = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.lif_vReset      = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.lif_gain      = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.lif_bias      = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.current    = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.extCurrent = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.totalCurrent = new float[managerRTDSize.maxNumNReg];
    managerRuntimeData.curSpike   = new bool[managerRTDSize.maxNumNReg];
    memset(managerRuntimeData.voltage, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.nextVoltage, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.recovery, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.Izh_a, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.Izh_b, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.Izh_c, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.Izh_d, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.Izh_C, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.Izh_k, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.Izh_vr, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.Izh_vt, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.Izh_vpeak, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.lif_tau_m, 0, sizeof(int) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.lif_tau_ref, 0, sizeof(int) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.lif_tau_ref_c, 0, sizeof(int) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.lif_vTh, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.lif_vReset, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.lif_gain, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.lif_bias, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.current, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.extCurrent, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.totalCurrent, 0, sizeof(float) * managerRTDSize.maxNumNReg);
    memset(managerRuntimeData.curSpike, 0, sizeof(bool) * managerRTDSize.maxNumNReg);
 
    managerRuntimeData.nVBuffer = new float[MAX_NEURON_MON_GRP_SZIE * 1000 * managerRTDSize.maxNumGroups]; // 1 second v buffer
    managerRuntimeData.nUBuffer = new float[MAX_NEURON_MON_GRP_SZIE * 1000 * managerRTDSize.maxNumGroups];
    managerRuntimeData.nIBuffer = new float[MAX_NEURON_MON_GRP_SZIE * 1000 * managerRTDSize.maxNumGroups];
    memset(managerRuntimeData.nVBuffer, 0, sizeof(float) * MAX_NEURON_MON_GRP_SZIE * 1000 * managerRTDSize.maxNumGroups);
    memset(managerRuntimeData.nUBuffer, 0, sizeof(float) * MAX_NEURON_MON_GRP_SZIE * 1000 * managerRTDSize.maxNumGroups);
    memset(managerRuntimeData.nIBuffer, 0, sizeof(float) * MAX_NEURON_MON_GRP_SZIE * 1000 * managerRTDSize.maxNumGroups);
 
    managerRuntimeData.gAMPA  = new float[managerRTDSize.glbNumNReg]; // sufficient to hold all regular neurons in the global network
    managerRuntimeData.gNMDA_r = new float[managerRTDSize.glbNumNReg]; // sufficient to hold all regular neurons in the global network
    managerRuntimeData.gNMDA_d = new float[managerRTDSize.glbNumNReg]; // sufficient to hold all regular neurons in the global network
    managerRuntimeData.gNMDA = new float[managerRTDSize.glbNumNReg]; // sufficient to hold all regular neurons in the global network
    memset(managerRuntimeData.gAMPA, 0, sizeof(float) * managerRTDSize.glbNumNReg);
    memset(managerRuntimeData.gNMDA_r, 0, sizeof(float) * managerRTDSize.glbNumNReg);
    memset(managerRuntimeData.gNMDA_d, 0, sizeof(float) * managerRTDSize.glbNumNReg);
    memset(managerRuntimeData.gNMDA, 0, sizeof(float) * managerRTDSize.glbNumNReg);
 
    managerRuntimeData.gGABAa = new float[managerRTDSize.glbNumNReg]; // sufficient to hold all regular neurons in the global network
    managerRuntimeData.gGABAb_r = new float[managerRTDSize.glbNumNReg]; // sufficient to hold all regular neurons in the global network
    managerRuntimeData.gGABAb_d = new float[managerRTDSize.glbNumNReg]; // sufficient to hold all regular neurons in the global network
    managerRuntimeData.gGABAb = new float[managerRTDSize.glbNumNReg]; // sufficient to hold all regular neurons in the global network
    memset(managerRuntimeData.gGABAa, 0, sizeof(float) * managerRTDSize.glbNumNReg);
    memset(managerRuntimeData.gGABAb_r, 0, sizeof(float) * managerRTDSize.glbNumNReg);
    memset(managerRuntimeData.gGABAb_d, 0, sizeof(float) * managerRTDSize.glbNumNReg);
    memset(managerRuntimeData.gGABAb, 0, sizeof(float) * managerRTDSize.glbNumNReg);
 
    // allocate neuromodulators and their assistive buffers
    managerRuntimeData.grpDA  = new float[managerRTDSize.maxNumGroups];
    managerRuntimeData.grp5HT = new float[managerRTDSize.maxNumGroups];
    managerRuntimeData.grpACh = new float[managerRTDSize.maxNumGroups];
    managerRuntimeData.grpNE  = new float[managerRTDSize.maxNumGroups];
    memset(managerRuntimeData.grpDA, 0, sizeof(float) * managerRTDSize.maxNumGroups);
    memset(managerRuntimeData.grp5HT, 0, sizeof(float) * managerRTDSize.maxNumGroups);
    memset(managerRuntimeData.grpACh, 0, sizeof(float) * managerRTDSize.maxNumGroups);
    memset(managerRuntimeData.grpNE, 0, sizeof(float) * managerRTDSize.maxNumGroups);
 
 
    managerRuntimeData.grpDABuffer  = new float[managerRTDSize.maxNumGroups * 1000]; // 1 second DA buffer
    managerRuntimeData.grp5HTBuffer = new float[managerRTDSize.maxNumGroups * 1000];
    managerRuntimeData.grpAChBuffer = new float[managerRTDSize.maxNumGroups * 1000];
    managerRuntimeData.grpNEBuffer  = new float[managerRTDSize.maxNumGroups * 1000];
    memset(managerRuntimeData.grpDABuffer, 0, managerRTDSize.maxNumGroups * sizeof(float) * 1000);
    memset(managerRuntimeData.grp5HTBuffer, 0, managerRTDSize.maxNumGroups * sizeof(float) * 1000);
    memset(managerRuntimeData.grpAChBuffer, 0, managerRTDSize.maxNumGroups * sizeof(float) * 1000);
    memset(managerRuntimeData.grpNEBuffer, 0, managerRTDSize.maxNumGroups * sizeof(float) * 1000);
 
    managerRuntimeData.lastSpikeTime = new int[managerRTDSize.maxNumNAssigned];
    memset(managerRuntimeData.lastSpikeTime, 0, sizeof(int) * managerRTDSize.maxNumNAssigned);
 
    managerRuntimeData.nSpikeCnt = new int[managerRTDSize.glbNumN];
    memset(managerRuntimeData.nSpikeCnt, 0, sizeof(int) * managerRTDSize.glbNumN); // sufficient to hold all neurons in the global network
 
    managerRuntimeData.avgFiring  = new float[managerRTDSize.maxNumN];
    managerRuntimeData.baseFiring = new float[managerRTDSize.maxNumN];
    memset(managerRuntimeData.avgFiring, 0, sizeof(float) * managerRTDSize.maxNumN);
    memset(managerRuntimeData.baseFiring, 0, sizeof(float) * managerRTDSize.maxNumN);
 
    // STP can be applied to spike generators, too -> numN
    // \TODO: The size of these data structures could be reduced to the max synaptic delay of all
    // connections with STP. That number might not be the same as maxDelay_.
    managerRuntimeData.stpu = new float[managerRTDSize.maxNumN * (glbNetworkConfig.maxDelay + 1)];
    managerRuntimeData.stpx = new float[managerRTDSize.maxNumN * (glbNetworkConfig.maxDelay + 1)];
    memset(managerRuntimeData.stpu, 0, sizeof(float) * managerRTDSize.maxNumN * (glbNetworkConfig.maxDelay + 1));
    memset(managerRuntimeData.stpx, 0, sizeof(float) * managerRTDSize.maxNumN * (glbNetworkConfig.maxDelay + 1));
 
    managerRuntimeData.Npre           = new unsigned short[managerRTDSize.maxNumNAssigned];
    managerRuntimeData.Npre_plastic   = new unsigned short[managerRTDSize.maxNumNAssigned];
    managerRuntimeData.Npost          = new unsigned short[managerRTDSize.maxNumNAssigned];
    managerRuntimeData.cumulativePost = new unsigned int[managerRTDSize.maxNumNAssigned];
    managerRuntimeData.cumulativePre  = new unsigned int[managerRTDSize.maxNumNAssigned];
    memset(managerRuntimeData.Npre, 0, sizeof(short) * managerRTDSize.maxNumNAssigned);
    memset(managerRuntimeData.Npre_plastic, 0, sizeof(short) * managerRTDSize.maxNumNAssigned);
    memset(managerRuntimeData.Npost, 0, sizeof(short) * managerRTDSize.maxNumNAssigned);
    memset(managerRuntimeData.cumulativePost, 0, sizeof(int) * managerRTDSize.maxNumNAssigned);
    memset(managerRuntimeData.cumulativePre, 0, sizeof(int) * managerRTDSize.maxNumNAssigned);
 
    managerRuntimeData.postSynapticIds = new SynInfo[managerRTDSize.maxNumPostSynNet];
    managerRuntimeData.postDelayInfo   = new DelayInfo[managerRTDSize.maxNumNAssigned * (glbNetworkConfig.maxDelay + 1)];   
    memset(managerRuntimeData.postSynapticIds, 0, sizeof(SynInfo) * managerRTDSize.maxNumPostSynNet);
    memset(managerRuntimeData.postDelayInfo, 0, sizeof(DelayInfo) * managerRTDSize.maxNumNAssigned * (glbNetworkConfig.maxDelay + 1));
 
    managerRuntimeData.preSynapticIds   = new SynInfo[managerRTDSize.maxNumPreSynNet];
    memset(managerRuntimeData.preSynapticIds, 0, sizeof(SynInfo) * managerRTDSize.maxNumPreSynNet);
 
    managerRuntimeData.wt           = new float[managerRTDSize.maxNumPreSynNet];
    managerRuntimeData.wtChange     = new float[managerRTDSize.maxNumPreSynNet];
    managerRuntimeData.maxSynWt     = new float[managerRTDSize.maxNumPreSynNet];
    managerRuntimeData.synSpikeTime = new int[managerRTDSize.maxNumPreSynNet];
    memset(managerRuntimeData.wt, 0, sizeof(float) * managerRTDSize.maxNumPreSynNet);
    memset(managerRuntimeData.wtChange, 0, sizeof(float) * managerRTDSize.maxNumPreSynNet);
    memset(managerRuntimeData.maxSynWt, 0, sizeof(float) * managerRTDSize.maxNumPreSynNet);
    memset(managerRuntimeData.synSpikeTime, 0, sizeof(int) * managerRTDSize.maxNumPreSynNet);
 
    mulSynFast = new float[managerRTDSize.maxNumConnections];
    mulSynSlow = new float[managerRTDSize.maxNumConnections];
    memset(mulSynFast, 0, sizeof(float) * managerRTDSize.maxNumConnections);
    memset(mulSynSlow, 0, sizeof(float) * managerRTDSize.maxNumConnections);
 
    managerRuntimeData.connIdsPreIdx    = new short int[managerRTDSize.maxNumPreSynNet];
    memset(managerRuntimeData.connIdsPreIdx, 0, sizeof(short int) * managerRTDSize.maxNumPreSynNet);
 
    managerRuntimeData.grpIds = new short int[managerRTDSize.maxNumNAssigned];
    memset(managerRuntimeData.grpIds, 0, sizeof(short int) * managerRTDSize.maxNumNAssigned);
 
    managerRuntimeData.spikeGenBits = new unsigned int[managerRTDSize.maxNumNSpikeGen / 32 + 1];
 
    // Confirm allocation of SNN runtime data in main memory
    managerRuntimeData.allocated = true;
    managerRuntimeData.memType = CPU_MEM;
}
 
int SNN::assignGroup(int gGrpId, int availableNeuronId) {
    int newAvailableNeuronId;
    assert(groupConfigMDMap[gGrpId].gStartN == -1); // The group has not yet been assigned
    groupConfigMDMap[gGrpId].gStartN = availableNeuronId;
    groupConfigMDMap[gGrpId].gEndN = availableNeuronId + groupConfigMap[gGrpId].numN - 1;
 
    KERNEL_DEBUG("Allocation for %d(%s), St=%d, End=%d",
                gGrpId, groupConfigMap[gGrpId].grpName.c_str(), groupConfigMDMap[gGrpId].gStartN, groupConfigMDMap[gGrpId].gEndN);
 
    newAvailableNeuronId = availableNeuronId + groupConfigMap[gGrpId].numN;
    //assert(newAvailableNeuronId <= numN);
 
    return newAvailableNeuronId;
}
 
int SNN::assignGroup(std::list<GroupConfigMD>::iterator grpIt, int localGroupId, int availableNeuronId) {
    int newAvailableNeuronId;
    assert(grpIt->lGrpId == -1); // The group has not yet been assigned
    grpIt->lGrpId = localGroupId;
    grpIt->lStartN = availableNeuronId;
    grpIt->lEndN = availableNeuronId + groupConfigMap[grpIt->gGrpId].numN - 1;
 
    grpIt->LtoGOffset = grpIt->gStartN - grpIt->lStartN;
    grpIt->GtoLOffset = grpIt->lStartN - grpIt->gStartN;
 
    KERNEL_DEBUG("Allocation for group (%s) [id:%d, local id:%d], St=%d, End=%d", groupConfigMap[grpIt->gGrpId].grpName.c_str(),
        grpIt->gGrpId, grpIt->lGrpId, grpIt->lStartN, grpIt->lEndN);
 
    newAvailableNeuronId = availableNeuronId + groupConfigMap[grpIt->gGrpId].numN;
 
    return newAvailableNeuronId;
}
 
void SNN::generateGroupRuntime(int netId, int lGrpId) {
    resetNeuromodulator(netId, lGrpId);
 
    for(int lNId = groupConfigs[netId][lGrpId].lStartN; lNId <= groupConfigs[netId][lGrpId].lEndN; lNId++)
        resetNeuron(netId, lGrpId, lNId);
}
 
void SNN::generateRuntimeGroupConfigs() {
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[netId].begin(); grpIt != groupPartitionLists[netId].end(); grpIt++) {
            // publish the group configs in an array for quick access and accessible on GPUs (cuda doesn't support std::list)
            int gGrpId = grpIt->gGrpId;
            int lGrpId = grpIt->lGrpId;
 
            // Data published by groupConfigMDMap[] are generated in compileSNN() and are invariant in partitionSNN()
            // Data published by grpIt are generated in partitionSNN() and maybe have duplicated copys
            groupConfigs[netId][lGrpId].netId = grpIt->netId;
            groupConfigs[netId][lGrpId].gGrpId = grpIt->gGrpId;
            groupConfigs[netId][lGrpId].gStartN = grpIt->gStartN;
            groupConfigs[netId][lGrpId].gEndN = grpIt->gEndN;
            groupConfigs[netId][lGrpId].lGrpId = grpIt->lGrpId;
            groupConfigs[netId][lGrpId].lStartN = grpIt->lStartN;
            groupConfigs[netId][lGrpId].lEndN = grpIt->lEndN;
            groupConfigs[netId][lGrpId].LtoGOffset = grpIt->LtoGOffset;
            groupConfigs[netId][lGrpId].GtoLOffset = grpIt->GtoLOffset;
            groupConfigs[netId][lGrpId].Type = groupConfigMap[gGrpId].type;
            groupConfigs[netId][lGrpId].numN = groupConfigMap[gGrpId].numN;
            groupConfigs[netId][lGrpId].numPostSynapses = grpIt->numPostSynapses;
            groupConfigs[netId][lGrpId].numPreSynapses = grpIt->numPreSynapses;
            groupConfigs[netId][lGrpId].isSpikeGenerator = groupConfigMap[gGrpId].isSpikeGenerator;
            groupConfigs[netId][lGrpId].isSpikeGenFunc = groupConfigMap[gGrpId].spikeGenFunc != NULL ? true : false;
            groupConfigs[netId][lGrpId].WithSTP =  groupConfigMap[gGrpId].stpConfig.WithSTP;
            groupConfigs[netId][lGrpId].WithSTDP =  groupConfigMap[gGrpId].stdpConfig.WithSTDP;
            groupConfigs[netId][lGrpId].WithESTDP =  groupConfigMap[gGrpId].stdpConfig.WithESTDP;
            groupConfigs[netId][lGrpId].WithISTDP = groupConfigMap[gGrpId].stdpConfig.WithISTDP;
            groupConfigs[netId][lGrpId].WithESTDPtype = groupConfigMap[gGrpId].stdpConfig.WithESTDPtype;
            groupConfigs[netId][lGrpId].WithISTDPtype =  groupConfigMap[gGrpId].stdpConfig.WithISTDPtype;
            groupConfigs[netId][lGrpId].WithESTDPcurve =  groupConfigMap[gGrpId].stdpConfig.WithESTDPcurve;
            groupConfigs[netId][lGrpId].WithISTDPcurve =  groupConfigMap[gGrpId].stdpConfig.WithISTDPcurve;
            groupConfigs[netId][lGrpId].WithHomeostasis =  groupConfigMap[gGrpId].homeoConfig.WithHomeostasis;
            groupConfigs[netId][lGrpId].FixedInputWts = grpIt->fixedInputWts;
            groupConfigs[netId][lGrpId].hasExternalConnect = grpIt->hasExternalConnect;
            groupConfigs[netId][lGrpId].Noffset = grpIt->Noffset; // Note: Noffset is not valid at this time
            groupConfigs[netId][lGrpId].MaxDelay = grpIt->maxOutgoingDelay;
            groupConfigs[netId][lGrpId].STP_A = groupConfigMap[gGrpId].stpConfig.STP_A;
            groupConfigs[netId][lGrpId].STP_U = groupConfigMap[gGrpId].stpConfig.STP_U;
            groupConfigs[netId][lGrpId].STP_tau_u_inv = groupConfigMap[gGrpId].stpConfig.STP_tau_u_inv;
            groupConfigs[netId][lGrpId].STP_tau_x_inv = groupConfigMap[gGrpId].stpConfig.STP_tau_x_inv;
            groupConfigs[netId][lGrpId].TAU_PLUS_INV_EXC = groupConfigMap[gGrpId].stdpConfig.TAU_PLUS_INV_EXC;
            groupConfigs[netId][lGrpId].TAU_MINUS_INV_EXC = groupConfigMap[gGrpId].stdpConfig.TAU_MINUS_INV_EXC;
            groupConfigs[netId][lGrpId].ALPHA_PLUS_EXC = groupConfigMap[gGrpId].stdpConfig.ALPHA_PLUS_EXC;
            groupConfigs[netId][lGrpId].ALPHA_MINUS_EXC = groupConfigMap[gGrpId].stdpConfig.ALPHA_MINUS_EXC;
            groupConfigs[netId][lGrpId].GAMMA = groupConfigMap[gGrpId].stdpConfig.GAMMA;
            groupConfigs[netId][lGrpId].KAPPA = groupConfigMap[gGrpId].stdpConfig.KAPPA;
            groupConfigs[netId][lGrpId].OMEGA = groupConfigMap[gGrpId].stdpConfig.OMEGA;
            groupConfigs[netId][lGrpId].TAU_PLUS_INV_INB = groupConfigMap[gGrpId].stdpConfig.TAU_PLUS_INV_INB;
            groupConfigs[netId][lGrpId].TAU_MINUS_INV_INB = groupConfigMap[gGrpId].stdpConfig.TAU_MINUS_INV_INB;
            groupConfigs[netId][lGrpId].ALPHA_PLUS_INB = groupConfigMap[gGrpId].stdpConfig.ALPHA_PLUS_INB;
            groupConfigs[netId][lGrpId].ALPHA_MINUS_INB = groupConfigMap[gGrpId].stdpConfig.ALPHA_MINUS_INB;
            groupConfigs[netId][lGrpId].BETA_LTP = groupConfigMap[gGrpId].stdpConfig.BETA_LTP;
            groupConfigs[netId][lGrpId].BETA_LTD = groupConfigMap[gGrpId].stdpConfig.BETA_LTD;
            groupConfigs[netId][lGrpId].LAMBDA = groupConfigMap[gGrpId].stdpConfig.LAMBDA;
            groupConfigs[netId][lGrpId].DELTA = groupConfigMap[gGrpId].stdpConfig.DELTA;
 
            groupConfigs[netId][lGrpId].numCompNeighbors = 0;
            groupConfigs[netId][lGrpId].withCompartments = groupConfigMap[gGrpId].withCompartments;
            groupConfigs[netId][lGrpId].compCouplingUp = groupConfigMap[gGrpId].compCouplingUp;
            groupConfigs[netId][lGrpId].compCouplingDown = groupConfigMap[gGrpId].compCouplingDown;
            memset(&groupConfigs[netId][lGrpId].compNeighbors, 0, sizeof(groupConfigs[netId][lGrpId].compNeighbors[0])*MAX_NUM_COMP_CONN);
            memset(&groupConfigs[netId][lGrpId].compCoupling, 0, sizeof(groupConfigs[netId][lGrpId].compCoupling[0])*MAX_NUM_COMP_CONN);
 
            groupConfigs[netId][lGrpId].avgTimeScale = groupConfigMap[gGrpId].homeoConfig.avgTimeScale;
            groupConfigs[netId][lGrpId].avgTimeScale_decay = groupConfigMap[gGrpId].homeoConfig.avgTimeScaleDecay;
            groupConfigs[netId][lGrpId].avgTimeScaleInv = groupConfigMap[gGrpId].homeoConfig.avgTimeScaleInv;
            groupConfigs[netId][lGrpId].homeostasisScale = groupConfigMap[gGrpId].homeoConfig.homeostasisScale;
 
            // parameters of neuromodulator
            groupConfigs[netId][lGrpId].baseDP = groupConfigMap[gGrpId].neuromodulatorConfig.baseDP;
            groupConfigs[netId][lGrpId].base5HT = groupConfigMap[gGrpId].neuromodulatorConfig.base5HT;
            groupConfigs[netId][lGrpId].baseACh = groupConfigMap[gGrpId].neuromodulatorConfig.baseACh;
            groupConfigs[netId][lGrpId].baseNE = groupConfigMap[gGrpId].neuromodulatorConfig.baseNE;
            groupConfigs[netId][lGrpId].decayDP = groupConfigMap[gGrpId].neuromodulatorConfig.decayDP;
            groupConfigs[netId][lGrpId].decay5HT = groupConfigMap[gGrpId].neuromodulatorConfig.decay5HT;
            groupConfigs[netId][lGrpId].decayACh = groupConfigMap[gGrpId].neuromodulatorConfig.decayACh;
            groupConfigs[netId][lGrpId].decayNE = groupConfigMap[gGrpId].neuromodulatorConfig.decayNE;
 
            // sync groupConfigs[][] and groupConfigMDMap[]
            if (netId == grpIt->netId) {
                groupConfigMDMap[gGrpId].netId = grpIt->netId;
                groupConfigMDMap[gGrpId].gGrpId = grpIt->gGrpId;
                groupConfigMDMap[gGrpId].gStartN = grpIt->gStartN;
                groupConfigMDMap[gGrpId].gEndN = grpIt->gEndN;
                groupConfigMDMap[gGrpId].lGrpId = grpIt->lGrpId;
                groupConfigMDMap[gGrpId].lStartN = grpIt->lStartN;
                groupConfigMDMap[gGrpId].lEndN = grpIt->lEndN;
                groupConfigMDMap[gGrpId].numPostSynapses = grpIt->numPostSynapses;
                groupConfigMDMap[gGrpId].numPreSynapses = grpIt->numPreSynapses;
                groupConfigMDMap[gGrpId].LtoGOffset = grpIt->LtoGOffset;
                groupConfigMDMap[gGrpId].GtoLOffset = grpIt->GtoLOffset;
                groupConfigMDMap[gGrpId].fixedInputWts = grpIt->fixedInputWts;
                groupConfigMDMap[gGrpId].hasExternalConnect = grpIt->hasExternalConnect;
                groupConfigMDMap[gGrpId].Noffset = grpIt->Noffset; // Note: Noffset is not valid at this time
                groupConfigMDMap[gGrpId].maxOutgoingDelay = grpIt->maxOutgoingDelay;
            }
            groupConfigs[netId][lGrpId].withParamModel_9 = groupConfigMap[gGrpId].withParamModel_9;
            groupConfigs[netId][lGrpId].isLIF = groupConfigMap[gGrpId].isLIF;
 
        }
 
        // FIXME: How does networkConfigs[netId].numGroups be availabe at this time?! Bug?!
        //int numNSpikeGen = 0;
        //for(int lGrpId = 0; lGrpId < networkConfigs[netId].numGroups; lGrpId++) {
        //  if (netId == groupConfigs[netId][lGrpId].netId && groupConfigs[netId][lGrpId].isSpikeGenerator && groupConfigs[netId][lGrpId].isSpikeGenFunc) {
        //  // we only need numNSpikeGen for spike generator callbacks that need to transfer their spikes to the GPU
        //      groupConfigs[netId][lGrpId].Noffset = numNSpikeGen; // FIXME, Noffset is updated after publish group configs
        //      numNSpikeGen += groupConfigs[netId][lGrpId].numN;
        //  }
        //}
        //assert(numNSpikeGen <= networkConfigs[netId].numNPois);
    }
}
 
void SNN::generateRuntimeConnectConfigs() {
    // sync localConnectLists and connectConfigMap
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        for (std::list<ConnectConfig>::iterator connIt = localConnectLists[netId].begin(); connIt != localConnectLists[netId].end(); connIt++) {
            connectConfigMap[connIt->connId] = *connIt;
        }
 
        for (std::list<ConnectConfig>::iterator connIt = externalConnectLists[netId].begin(); connIt != externalConnectLists[netId].end(); connIt++) {
            connectConfigMap[connIt->connId] = *connIt;
        }
    }
}
 
void SNN::generateRuntimeNetworkConfigs() {
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            // copy the global network config to local network configs
            // global configuration for maximum axonal delay
            networkConfigs[netId].maxDelay  = glbNetworkConfig.maxDelay;
 
            // configurations for execution features
            networkConfigs[netId].sim_with_fixedwts = sim_with_fixedwts;
            networkConfigs[netId].sim_with_conductances = sim_with_conductances;
            networkConfigs[netId].sim_with_homeostasis = sim_with_homeostasis;
            networkConfigs[netId].sim_with_stdp = sim_with_stdp;
            networkConfigs[netId].sim_with_stp = sim_with_stp;
            networkConfigs[netId].sim_in_testing = sim_in_testing;
 
            // search for active neuron monitor
            networkConfigs[netId].sim_with_nm = false;
            for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[netId].begin(); grpIt != groupPartitionLists[netId].end(); grpIt++) {
                if (grpIt->netId == netId && grpIt->neuronMonitorId >= 0)
                    networkConfigs[netId].sim_with_nm = true;
            }
 
            // stdp, da-stdp configurations
            networkConfigs[netId].stdpScaleFactor = stdpScaleFactor_;
            networkConfigs[netId].wtChangeDecay = wtChangeDecay_;
 
            // conductance configurations
            networkConfigs[netId].sim_with_NMDA_rise = sim_with_NMDA_rise;
            networkConfigs[netId].sim_with_GABAb_rise = sim_with_GABAb_rise;
            networkConfigs[netId].dAMPA = dAMPA;
            networkConfigs[netId].rNMDA = rNMDA;
            networkConfigs[netId].dNMDA = dNMDA;
            networkConfigs[netId].sNMDA = sNMDA;
            networkConfigs[netId].dGABAa = dGABAa;
            networkConfigs[netId].rGABAb = rGABAb;
            networkConfigs[netId].dGABAb = dGABAb;
            networkConfigs[netId].sGABAb = sGABAb;
 
            networkConfigs[netId].simIntegrationMethod = glbNetworkConfig.simIntegrationMethod;
            networkConfigs[netId].simNumStepsPerMs = glbNetworkConfig.simNumStepsPerMs;
            networkConfigs[netId].timeStep = glbNetworkConfig.timeStep;
 
            // configurations for boundries of neural types
            findNumN(netId, networkConfigs[netId].numN, networkConfigs[netId].numNExternal, networkConfigs[netId].numNAssigned,
                     networkConfigs[netId].numNReg, networkConfigs[netId].numNExcReg, networkConfigs[netId].numNInhReg,
                     networkConfigs[netId].numNPois, networkConfigs[netId].numNExcPois, networkConfigs[netId].numNInhPois);
 
            // configurations for assigned groups and connections
            networkConfigs[netId].numGroups = 0;
            for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[netId].begin(); grpIt != groupPartitionLists[netId].end(); grpIt++) {
                if (grpIt->netId == netId)
                    networkConfigs[netId].numGroups++;
            }
            networkConfigs[netId].numGroupsAssigned = groupPartitionLists[netId].size();
            //networkConfigs[netId].numConnections = localConnectLists[netId].size();
            //networkConfigs[netId].numAssignedConnections = localConnectLists[netId].size() + externalConnectLists[netId].size();
            //networkConfigs[netId].numConnections = localConnectLists[netId].size() + externalConnectLists[netId].size();
            networkConfigs[netId].numConnections = connectConfigMap.size();// temporarily solution: copy all connection info to each GPU
 
            // find the maximum number of pre- and post-connections among neurons
            // SNN::maxNumPreSynN and SNN::maxNumPostSynN are updated
            findMaxNumSynapsesNeurons(netId, networkConfigs[netId].maxNumPostSynN, networkConfigs[netId].maxNumPreSynN);
 
            // find the maximum number of spikes in D1 (i.e., maxDelay == 1) and D2 (i.e., maxDelay >= 2) sets
            findMaxSpikesD1D2(netId, networkConfigs[netId].maxSpikesD1, networkConfigs[netId].maxSpikesD2);
 
            // find the total number of synapses in the network
            findNumSynapsesNetwork(netId, networkConfigs[netId].numPostSynNet, networkConfigs[netId].numPreSynNet);
 
            // find out number of user-defined spike gen and update Noffset of each group config
            // Note: groupConfigs[][].Noffset is valid at this time
            findNumNSpikeGenAndOffset(netId);
        }
    }
 
    // find manager runtime data size, which is sufficient to hold the data of any gpu runtime
    memset(&managerRTDSize, 0, sizeof(ManagerRuntimeDataSize));
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            // find the maximum number of numN, numNReg ,and numNAssigned among local networks
            if (networkConfigs[netId].numNReg > managerRTDSize.maxNumNReg) managerRTDSize.maxNumNReg = networkConfigs[netId].numNReg;
            if (networkConfigs[netId].numN > managerRTDSize.maxNumN) managerRTDSize.maxNumN = networkConfigs[netId].numN;
            if (networkConfigs[netId].numNAssigned > managerRTDSize.maxNumNAssigned) managerRTDSize.maxNumNAssigned = networkConfigs[netId].numNAssigned;
 
            // find the maximum number of numNSpikeGen among local networks
            if (networkConfigs[netId].numNSpikeGen > managerRTDSize.maxNumNSpikeGen) managerRTDSize.maxNumNSpikeGen = networkConfigs[netId].numNSpikeGen;
 
            // find the maximum number of numGroups and numConnections among local networks
            if (networkConfigs[netId].numGroups > managerRTDSize.maxNumGroups) managerRTDSize.maxNumGroups = networkConfigs[netId].numGroups;
            if (networkConfigs[netId].numConnections > managerRTDSize.maxNumConnections) managerRTDSize.maxNumConnections = networkConfigs[netId].numConnections;
 
            // find the maximum number of neurons in a group among local networks
            for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[netId].begin(); grpIt != groupPartitionLists[netId].end(); grpIt++) {
                if (groupConfigMap[grpIt->gGrpId].numN > managerRTDSize.maxNumNPerGroup) managerRTDSize.maxNumNPerGroup = groupConfigMap[grpIt->gGrpId].numN;
            }
 
            // find the maximum number of maxSipkesD1(D2) among networks
            if (networkConfigs[netId].maxSpikesD1 > managerRTDSize.maxMaxSpikeD1) managerRTDSize.maxMaxSpikeD1 = networkConfigs[netId].maxSpikesD1;
            if (networkConfigs[netId].maxSpikesD2 > managerRTDSize.maxMaxSpikeD2) managerRTDSize.maxMaxSpikeD2 = networkConfigs[netId].maxSpikesD2;
 
            // find the maximum number of total # of pre- and post-connections among local networks
            if (networkConfigs[netId].numPreSynNet > managerRTDSize.maxNumPreSynNet) managerRTDSize.maxNumPreSynNet = networkConfigs[netId].numPreSynNet;
            if (networkConfigs[netId].numPostSynNet > managerRTDSize.maxNumPostSynNet) managerRTDSize.maxNumPostSynNet = networkConfigs[netId].numPostSynNet;
 
            // find the number of numN, and numNReg in the global network
            managerRTDSize.glbNumN += networkConfigs[netId].numN;
            managerRTDSize.glbNumNReg += networkConfigs[netId].numNReg;
        }
    }
}
 
bool compareSrcNeuron(const ConnectionInfo& first, const ConnectionInfo& second) {
    return (first.nSrc + first.srcGLoffset < second.nSrc + second.srcGLoffset);
}
 
bool compareDelay(const ConnectionInfo& first, const ConnectionInfo& second) {
    return (first.delay < second.delay);
}
 
// Note: ConnectInfo stored in connectionList use global ids
void SNN::generateConnectionRuntime(int netId) {
    std::map<int, int> GLoffset; // global nId to local nId offset
    std::map<int, int> GLgrpId; // global grpId to local grpId offset
 
    // load offset between global neuron id and local neuron id
    for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[netId].begin(); grpIt != groupPartitionLists[netId].end(); grpIt++) {
        GLoffset[grpIt->gGrpId] = grpIt->GtoLOffset;
        GLgrpId[grpIt->gGrpId] = grpIt->lGrpId;
    }
    // FIXME: connId is global connId, use connectConfigs[netId][local connId] instead,
    // FIXME; but note connectConfigs[netId][] are NOT complete, lack of exeternal incoming connections
    // generate mulSynFast, mulSynSlow in connection-centric array
    for (std::map<int, ConnectConfig>::iterator connIt = connectConfigMap.begin(); connIt != connectConfigMap.end(); connIt++) {
        // store scaling factors for synaptic currents in connection-centric array
        mulSynFast[connIt->second.connId] = connIt->second.mulSynFast;
        mulSynSlow[connIt->second.connId] = connIt->second.mulSynSlow;
    }
 
    // parse ConnectionInfo stored in connectionLists[0]
    // note: ConnectInfo stored in connectionList use global ids
    // generate Npost, Npre, Npre_plastic
    int parsedConnections = 0;
    memset(managerRuntimeData.Npost, 0, sizeof(short) * networkConfigs[netId].numNAssigned);
    memset(managerRuntimeData.Npre, 0, sizeof(short) * networkConfigs[netId].numNAssigned);
    memset(managerRuntimeData.Npre_plastic, 0, sizeof(short) * networkConfigs[netId].numNAssigned);
    for (std::list<ConnectionInfo>::iterator connIt = connectionLists[netId].begin(); connIt != connectionLists[netId].end(); connIt++) {
        connIt->srcGLoffset = GLoffset[connIt->grpSrc];
        if (managerRuntimeData.Npost[connIt->nSrc + GLoffset[connIt->grpSrc]] == SYNAPSE_ID_MASK) {
            KERNEL_ERROR("Error: the number of synapses exceeds maximum limit (%d) for neuron %d (group %d)", SYNAPSE_ID_MASK, connIt->nSrc, connIt->grpSrc);
            exitSimulation(ID_OVERFLOW_ERROR);
        }
        if (managerRuntimeData.Npre[connIt->nDest + GLoffset[connIt->grpDest]] == SYNAPSE_ID_MASK) {
            KERNEL_ERROR("Error: the number of synapses exceeds maximum limit (%d) for neuron %d (group %d)", SYNAPSE_ID_MASK, connIt->nDest, connIt->grpDest);
            exitSimulation(ID_OVERFLOW_ERROR);
        }
        managerRuntimeData.Npost[connIt->nSrc + GLoffset[connIt->grpSrc]]++;
        managerRuntimeData.Npre[connIt->nDest + GLoffset[connIt->grpDest]]++;
 
        if (GET_FIXED_PLASTIC(connectConfigMap[connIt->connId].connProp) == SYN_PLASTIC) {
            sim_with_fixedwts = false; // if network has any plastic synapses at all, this will be set to true
            managerRuntimeData.Npre_plastic[connIt->nDest + GLoffset[connIt->grpDest]]++;
 
            // homeostasis
            if (groupConfigMap[connIt->grpDest].homeoConfig.WithHomeostasis && groupConfigMDMap[connIt->grpDest].homeoId == -1)
                groupConfigMDMap[connIt->grpDest].homeoId = connIt->nDest + GLoffset[connIt->grpDest]; // this neuron info will be printed
 
            // old access to homeostasis
            //if (groupConfigs[netId][GLgrpId[it->grpDest]].WithHomeostasis && groupConfigs[netId][GLgrpId[it->grpDest]].homeoId == -1)
            //  groupConfigs[netId][GLgrpId[it->grpDest]].homeoId = it->nDest + GLoffset[it->grpDest]; // this neuron info will be printed
        }
 
        // generate the delay vaule
        //it->delay = connectConfigMap[it->connId].minDelay + rand() % (connectConfigMap[it->connId].maxDelay - connectConfigMap[it->connId].minDelay + 1);
        //assert((it->delay >= connectConfigMap[it->connId].minDelay) && (it->delay <= connectConfigMap[it->connId].maxDelay));
        // generate the max weight and initial weight
        //float initWt = generateWeight(connectConfigMap[it->connId].connProp, connectConfigMap[it->connId].initWt, connectConfigMap[it->connId].maxWt, it->nSrc, it->grpSrc);
        //float initWt = connectConfigMap[it->connId].initWt;
        //float maxWt = connectConfigMap[it->connId].maxWt;
        // adjust sign of weight based on pre-group (negative if pre is inhibitory)
        // this access is fine, isExcitatoryGroup() use global grpId
        //it->maxWt = isExcitatoryGroup(it->grpSrc) ? fabs(maxWt) : -1.0 * fabs(maxWt);
        //it->initWt = isExcitatoryGroup(it->grpSrc) ? fabs(initWt) : -1.0 * fabs(initWt);
 
        parsedConnections++;
    }
    assert(parsedConnections == networkConfigs[netId].numPostSynNet && parsedConnections == networkConfigs[netId].numPreSynNet);
 
    // generate cumulativePost and cumulativePre
    managerRuntimeData.cumulativePost[0] = 0;
    managerRuntimeData.cumulativePre[0] = 0;
    for (int lNId = 1; lNId < networkConfigs[netId].numNAssigned; lNId++) {
        managerRuntimeData.cumulativePost[lNId] = managerRuntimeData.cumulativePost[lNId - 1] + managerRuntimeData.Npost[lNId - 1];
        managerRuntimeData.cumulativePre[lNId] = managerRuntimeData.cumulativePre[lNId - 1] + managerRuntimeData.Npre[lNId - 1];
    }
 
    // generate preSynapticIds, parse plastic connections first
    memset(managerRuntimeData.Npre, 0, sizeof(short) * networkConfigs[netId].numNAssigned); // reset managerRuntimeData.Npre to zero, so that it can be used as synId
    parsedConnections = 0;
    for (std::list<ConnectionInfo>::iterator connIt = connectionLists[netId].begin(); connIt != connectionLists[netId].end(); connIt++) {
        if (GET_FIXED_PLASTIC(connectConfigMap[connIt->connId].connProp) == SYN_PLASTIC) {
            int pre_pos = managerRuntimeData.cumulativePre[connIt->nDest + GLoffset[connIt->grpDest]] + managerRuntimeData.Npre[connIt->nDest + GLoffset[connIt->grpDest]];
            assert(pre_pos < networkConfigs[netId].numPreSynNet);
 
            managerRuntimeData.preSynapticIds[pre_pos] = SET_CONN_ID((connIt->nSrc + GLoffset[connIt->grpSrc]), 0, (GLgrpId[connIt->grpSrc])); // managerRuntimeData.Npost[it->nSrc] is not availabe at this parse
            connIt->preSynId = managerRuntimeData.Npre[connIt->nDest + GLoffset[connIt->grpDest]]; // save managerRuntimeData.Npre[it->nDest] as synId
 
            managerRuntimeData.Npre[connIt->nDest+ GLoffset[connIt->grpDest]]++;
            parsedConnections++;
 
            // update the maximum number of and pre-connections of a neuron in a group
            //if (managerRuntimeData.Npre[it->nDest] > groupInfo[it->grpDest].maxPreConn)
            //  groupInfo[it->grpDest].maxPreConn = managerRuntimeData.Npre[it->nDest];
        }
    }
    // parse fixed connections
    for (std::list<ConnectionInfo>::iterator connIt = connectionLists[netId].begin(); connIt != connectionLists[netId].end(); connIt++) {
        if (GET_FIXED_PLASTIC(connectConfigMap[connIt->connId].connProp) == SYN_FIXED) {
            int pre_pos = managerRuntimeData.cumulativePre[connIt->nDest + GLoffset[connIt->grpDest]] + managerRuntimeData.Npre[connIt->nDest + GLoffset[connIt->grpDest]];
            assert(pre_pos < networkConfigs[netId].numPreSynNet);
 
            managerRuntimeData.preSynapticIds[pre_pos] = SET_CONN_ID((connIt->nSrc + GLoffset[connIt->grpSrc]), 0, (GLgrpId[connIt->grpSrc])); // managerRuntimeData.Npost[it->nSrc] is not availabe at this parse
            connIt->preSynId = managerRuntimeData.Npre[connIt->nDest + GLoffset[connIt->grpDest]]; // save managerRuntimeData.Npre[it->nDest] as synId
 
            managerRuntimeData.Npre[connIt->nDest + GLoffset[connIt->grpDest]]++;
            parsedConnections++;
 
            // update the maximum number of and pre-connections of a neuron in a group
            //if (managerRuntimeData.Npre[it->nDest] > groupInfo[it->grpDest].maxPreConn)
            //  groupInfo[it->grpDest].maxPreConn = managerRuntimeData.Npre[it->nDest];
        }
    }
    assert(parsedConnections == networkConfigs[netId].numPreSynNet);
    //printf("parsed pre connections %d\n", parsedConnections);
 
    // generate postSynapticIds
    connectionLists[netId].sort(compareSrcNeuron); // sort by local nSrc id
    memset(managerRuntimeData.postDelayInfo, 0, sizeof(DelayInfo) * (networkConfigs[netId].numNAssigned * (glbNetworkConfig.maxDelay + 1)));
    for (int lNId = 0; lNId < networkConfigs[netId].numNAssigned; lNId++) { // pre-neuron order, local nId
        if (managerRuntimeData.Npost[lNId] > 0) {
            std::list<ConnectionInfo> postConnectionList;
            ConnectionInfo targetConn;
            targetConn.nSrc = lNId ; // the other fields does not matter, use local nid to search
 
            std::list<ConnectionInfo>::iterator firstPostConn = std::find(connectionLists[netId].begin(), connectionLists[netId].end(), targetConn);
            std::list<ConnectionInfo>::iterator lastPostConn = firstPostConn;
            std::advance(lastPostConn, managerRuntimeData.Npost[lNId]);
            managerRuntimeData.Npost[lNId] = 0; // reset managerRuntimeData.Npost[lNId] to zero, so that it can be used as synId
 
            postConnectionList.splice(postConnectionList.begin(), connectionLists[netId], firstPostConn, lastPostConn);
            postConnectionList.sort(compareDelay);
 
            int post_pos, pre_pos, lastDelay = 0;
            parsedConnections = 0;
            //memset(&managerRuntimeData.postDelayInfo[lNId * (glbNetworkConfig.maxDelay + 1)], 0, sizeof(DelayInfo) * (glbNetworkConfig.maxDelay + 1));
            for (std::list<ConnectionInfo>::iterator connIt = postConnectionList.begin(); connIt != postConnectionList.end(); connIt++) {
                assert(connIt->nSrc + GLoffset[connIt->grpSrc] == lNId);
                post_pos = managerRuntimeData.cumulativePost[connIt->nSrc + GLoffset[connIt->grpSrc]] + managerRuntimeData.Npost[connIt->nSrc + GLoffset[connIt->grpSrc]];
                pre_pos  = managerRuntimeData.cumulativePre[connIt->nDest + GLoffset[connIt->grpDest]] + connIt->preSynId;
 
                assert(post_pos < networkConfigs[netId].numPostSynNet);
                //assert(pre_pos  < numPreSynNet);
 
                // generate a post synaptic id for the current connection
                managerRuntimeData.postSynapticIds[post_pos] = SET_CONN_ID((connIt->nDest + GLoffset[connIt->grpDest]), connIt->preSynId, (GLgrpId[connIt->grpDest]));// used stored managerRuntimeData.Npre[it->nDest] in it->preSynId
                // generate a delay look up table by the way
                assert(connIt->delay > 0);
                if (connIt->delay > lastDelay) {
                    managerRuntimeData.postDelayInfo[lNId * (glbNetworkConfig.maxDelay + 1) + connIt->delay - 1].delay_index_start = managerRuntimeData.Npost[connIt->nSrc + GLoffset[connIt->grpSrc]];
                    managerRuntimeData.postDelayInfo[lNId * (glbNetworkConfig.maxDelay + 1) + connIt->delay - 1].delay_length++;
                } else if (connIt->delay == lastDelay) {
                    managerRuntimeData.postDelayInfo[lNId * (glbNetworkConfig.maxDelay + 1) + connIt->delay - 1].delay_length++;
                } else {
                    KERNEL_ERROR("Post-synaptic delays not sorted correctly... pre_id=%d, delay[%d]=%d, delay[%d]=%d",
                        lNId, managerRuntimeData.Npost[connIt->nSrc + GLoffset[connIt->grpSrc]], connIt->delay, managerRuntimeData.Npost[connIt->nSrc + GLoffset[connIt->grpSrc]] - 1, lastDelay);
                }
                lastDelay = connIt->delay;
 
                // update the corresponding pre synaptic id
                SynInfo preId = managerRuntimeData.preSynapticIds[pre_pos];
                assert(GET_CONN_NEURON_ID(preId) == connIt->nSrc + GLoffset[connIt->grpSrc]);
                //assert(GET_CONN_GRP_ID(preId) == it->grpSrc);
                managerRuntimeData.preSynapticIds[pre_pos] = SET_CONN_ID((connIt->nSrc + GLoffset[connIt->grpSrc]), managerRuntimeData.Npost[connIt->nSrc + GLoffset[connIt->grpSrc]], (GLgrpId[connIt->grpSrc]));
                managerRuntimeData.wt[pre_pos] = connIt->initWt;
                managerRuntimeData.maxSynWt[pre_pos] = connIt->maxWt;
                managerRuntimeData.connIdsPreIdx[pre_pos] = connIt->connId;
 
                managerRuntimeData.Npost[connIt->nSrc + GLoffset[connIt->grpSrc]]++;
                parsedConnections++;
 
                // update the maximum number of and post-connections of a neuron in a group
                //if (managerRuntimeData.Npost[it->nSrc] > groupInfo[it->grpSrc].maxPostConn)
                //  groupInfo[it->grpSrc].maxPostConn = managerRuntimeData.Npost[it->nSrc];
            }
            assert(parsedConnections == managerRuntimeData.Npost[lNId]);
            //printf("parsed post connections %d\n", parsedConnections);
            // note: elements in postConnectionList are deallocated automatically with postConnectionList
            /* for postDelayInfo debugging
            printf("%d ", lNId);
            for (int t = 0; t < maxDelay_ + 1; t ++) {
                printf("[%d,%d]",
                    managerRuntimeData.postDelayInfo[lNId * (maxDelay_ + 1) + t].delay_index_start,
                    managerRuntimeData.postDelayInfo[lNId * (maxDelay_ + 1) + t].delay_length);
            }
            printf("\n");
            */
        }
    }
    assert(connectionLists[netId].empty());
 
    //int p = managerRuntimeData.Npost[src];
 
    //assert(managerRuntimeData.Npost[src] >= 0);
    //assert(managerRuntimeData.Npre[dest] >= 0);
    //assert((src * maxNumPostSynGrp + p) / numN < maxNumPostSynGrp); // divide by numN to prevent INT overflow
 
    //unsigned int post_pos = managerRuntimeData.cumulativePost[src] + managerRuntimeData.Npost[src];
    //unsigned int pre_pos  = managerRuntimeData.cumulativePre[dest] + managerRuntimeData.Npre[dest];
 
    //assert(post_pos < numPostSynNet);
    //assert(pre_pos  < numPreSynNet);
 
    //managerRuntimeData.postSynapticIds[post_pos]   = SET_CONN_ID(dest, managerRuntimeData.Npre[dest], destGrp);
    //tmp_SynapticDelay[post_pos] = dVal;
 
    //managerRuntimeData.preSynapticIds[pre_pos] = SET_CONN_ID(src, managerRuntimeData.Npost[src], srcGrp);
    //managerRuntimeData.wt[pre_pos]      = synWt;
    //managerRuntimeData.maxSynWt[pre_pos] = maxWt;
    //managerRuntimeData.connIdsPreIdx[pre_pos] = connId;
 
    //bool synWtType = GET_FIXED_PLASTIC(connProp);
 
    //if (synWtType == SYN_PLASTIC) {
    //  sim_with_fixedwts = false; // if network has any plastic synapses at all, this will be set to true
    //  managerRuntimeData.Npre_plastic[dest]++;
    //  // homeostasis
    //  if (groupConfigs[0][destGrp].WithHomeostasis && groupConfigs[0][destGrp].homeoId ==-1)
    //      groupConfigs[0][destGrp].homeoId = dest; // this neuron info will be printed
    //}
 
    //managerRuntimeData.Npre[dest] += 1;
    //managerRuntimeData.Npost[src] += 1;
 
    //groupInfo[srcGrp].numPostConn++;
    //groupInfo[destGrp].numPreConn++;
 
    //if (managerRuntimeData.Npost[src] > groupInfo[srcGrp].maxPostConn)
    //  groupInfo[srcGrp].maxPostConn = managerRuntimeData.Npost[src];
    //if (managerRuntimeData.Npre[dest] > groupInfo[destGrp].maxPreConn)
    //  groupInfo[destGrp].maxPreConn = managerRuntimeData.Npre[src];
}
 
void SNN::generateCompConnectionRuntime(int netId)
{
    std::map<int, int> GLgrpId; // global grpId to local grpId offset
 
    for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[netId].begin(); grpIt != groupPartitionLists[netId].end(); grpIt++) {
        GLgrpId[grpIt->gGrpId] = grpIt->lGrpId;
        //printf("Global group id %i; Local group id %i\n", grpIt->gGrpId, grpIt->lGrpId);
    }
 
    //printf("The current netid is: %i\n", netId);
 
    for (std::list<compConnectConfig>::iterator connIt = localCompConnectLists[netId].begin(); connIt != localCompConnectLists[netId].end(); connIt++) {
        //printf("The size of localCompConnectLists is: %i\n", localCompConnectLists[netId].size());
        int grpLower = connIt->grpSrc;
        int grpUpper = connIt->grpDest;
 
        int i = groupConfigs[netId][GLgrpId[grpLower]].numCompNeighbors;
        if (i >= MAX_NUM_COMP_CONN) {
            KERNEL_ERROR("Group %s(%d) exceeds max number of allowed compartmental connections (%d).",
                groupConfigMap[grpLower].grpName.c_str(), grpLower, (int)MAX_NUM_COMP_CONN);
            exitSimulation(1);
        }
        groupConfigs[netId][GLgrpId[grpLower]].compNeighbors[i] = grpUpper;
        groupConfigs[netId][GLgrpId[grpLower]].compCoupling[i] = groupConfigs[netId][GLgrpId[grpUpper]].compCouplingDown; // get down-coupling from upper neighbor
        groupConfigs[netId][GLgrpId[grpLower]].numCompNeighbors++;
 
        int j = groupConfigs[netId][GLgrpId[grpUpper]].numCompNeighbors;
        if (j >= MAX_NUM_COMP_CONN) {
            KERNEL_ERROR("Group %s(%d) exceeds max number of allowed compartmental connections (%d).",
                groupConfigMap[grpUpper].grpName.c_str(), grpUpper, (int)MAX_NUM_COMP_CONN);
            exitSimulation(1);
        }
        groupConfigs[netId][GLgrpId[grpUpper]].compNeighbors[j] = grpLower;
        groupConfigs[netId][GLgrpId[grpUpper]].compCoupling[j] = groupConfigs[netId][GLgrpId[grpLower]].compCouplingUp; // get up-coupling from lower neighbor
        groupConfigs[netId][GLgrpId[grpUpper]].numCompNeighbors++;
 
        //printf("Group %i (local group %i) has %i compartmental neighbors!\n", grpUpper, GLgrpId[grpUpper], groupConfigs[netId][GLgrpId[grpUpper]].numCompNeighbors);
    }
}
 
 
void SNN::generatePoissonGroupRuntime(int netId, int lGrpId) {
    for(int lNId = groupConfigs[netId][lGrpId].lStartN; lNId <= groupConfigs[netId][lGrpId].lEndN; lNId++)
        resetPoissonNeuron(netId, lGrpId, lNId);
}
 
 
void SNN::collectGlobalNetworkConfigC() {
    // scan all connect configs to find the maximum delay in the global network, update glbNetworkConfig.maxDelay
    for (std::map<int, ConnectConfig>::iterator connIt = connectConfigMap.begin(); connIt != connectConfigMap.end(); connIt++) {
        if (connIt->second.maxDelay > glbNetworkConfig.maxDelay)
            glbNetworkConfig.maxDelay = connIt->second.maxDelay;
    }
    assert(connectConfigMap.size() > 0 || glbNetworkConfig.maxDelay != -1);
 
    // scan all group configs to find the number of (reg, pois, exc, inh) neuron in the global network
    for(int gGrpId = 0; gGrpId < numGroups; gGrpId++) {
        if (IS_EXCITATORY_TYPE(groupConfigMap[gGrpId].type) && (groupConfigMap[gGrpId].type & POISSON_NEURON)) {
            glbNetworkConfig.numNExcPois += groupConfigMap[gGrpId].numN;
        } else if (IS_INHIBITORY_TYPE(groupConfigMap[gGrpId].type) &&  (groupConfigMap[gGrpId].type & POISSON_NEURON)) {
            glbNetworkConfig.numNInhPois += groupConfigMap[gGrpId].numN;
        } else if (IS_EXCITATORY_TYPE(groupConfigMap[gGrpId].type) && !(groupConfigMap[gGrpId].type & POISSON_NEURON)) {
            glbNetworkConfig.numNExcReg += groupConfigMap[gGrpId].numN;
        } else if (IS_INHIBITORY_TYPE(groupConfigMap[gGrpId].type) && !(groupConfigMap[gGrpId].type & POISSON_NEURON)) {
            glbNetworkConfig.numNInhReg += groupConfigMap[gGrpId].numN;
        }
 
        if (groupConfigMDMap[gGrpId].maxOutgoingDelay == 1)
            glbNetworkConfig.numN1msDelay += groupConfigMap[gGrpId].numN;
        else if (groupConfigMDMap[gGrpId].maxOutgoingDelay >= 2)
            glbNetworkConfig.numN2msDelay += groupConfigMap[gGrpId].numN;
    }
 
    glbNetworkConfig.numNReg = glbNetworkConfig.numNExcReg + glbNetworkConfig.numNInhReg;
    glbNetworkConfig.numNPois = glbNetworkConfig.numNExcPois + glbNetworkConfig.numNInhPois;
    glbNetworkConfig.numN = glbNetworkConfig.numNReg + glbNetworkConfig.numNPois;
}
 
 
void SNN::collectGlobalNetworkConfigP() {
    // print group and connection overview
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!localConnectLists[netId].empty() || !externalConnectLists[netId].empty()) {
            for (std::list<ConnectConfig>::iterator connIt = localConnectLists[netId].begin(); connIt != localConnectLists[netId].end(); connIt++)
                glbNetworkConfig.numSynNet += connIt->numberOfConnections;
 
            for (std::list<ConnectConfig>::iterator connIt = externalConnectLists[netId].begin(); connIt != externalConnectLists[netId].end(); connIt++)
                glbNetworkConfig.numSynNet += connIt->numberOfConnections;
        }
    }
}
 
// after all the initalization. Its time to create the synaptic weights, weight change and also
// time of firing these are the mostly costly arrays so dense packing is essential to minimize wastage of space
void SNN::compileSNN() {
    KERNEL_DEBUG("Beginning compilation of the network....");
 
    // compile (update) group and connection configs according to their mutual information
    // update GroupConfig::MaxDelay GroupConfig::FixedInputWts
    // assign GroupConfig::StartN and GroupConfig::EndN
    // Note: MaxDelay, FixedInputWts, StartN, and EndN are invariant in single-GPU or multi-GPUs mode
    compileGroupConfig();
 
    compileConnectConfig(); // for future use
 
    // collect the global network config according to compiled gorup and connection configs
    // collect SNN::maxDelay_
    collectGlobalNetworkConfigC();
 
    // perform various consistency checks:
    // - numNeurons vs. sum of all neurons
    // - STDP set on a post-group with incoming plastic connections
    // - etc.
    verifyNetwork();
 
    // display the global network configuration
    KERNEL_INFO("\n");
    KERNEL_INFO("************************** Global Network Configuration *******************************");
    KERNEL_INFO("The number of neurons in the network (numN) = %d", glbNetworkConfig.numN);
    KERNEL_INFO("The number of regular neurons in the network (numNReg:numNExcReg:numNInhReg) = %d:%d:%d", glbNetworkConfig.numNReg, glbNetworkConfig.numNExcReg, glbNetworkConfig.numNInhReg);
    KERNEL_INFO("The number of poisson neurons in the network (numNPois:numNExcPois:numInhPois) = %d:%d:%d", glbNetworkConfig.numNPois, glbNetworkConfig.numNExcPois, glbNetworkConfig.numNInhPois);
    KERNEL_INFO("The maximum axonal delay in the network (maxDelay) = %d", glbNetworkConfig.maxDelay);
 
    //ensure that we dont compile the network again
    snnState = COMPILED_SNN;
}
 
void SNN::compileConnectConfig() {
    // for future  use
}
 
void SNN::compileGroupConfig() {
    int grpSrc;
    bool synWtType;
 
    // find the maximum delay for each group according to incoming connection
    for (std::map<int, ConnectConfig>::iterator connIt = connectConfigMap.begin(); connIt != connectConfigMap.end(); connIt++) {
        // check if the current connection's delay meaning grpSrc's delay
        // is greater than the MaxDelay for grpSrc. We find the maximum
        // delay for the grpSrc by this scheme.
        grpSrc = connIt->second.grpSrc;
        if (connIt->second.maxDelay > groupConfigMDMap[grpSrc].maxOutgoingDelay)
            groupConfigMDMap[grpSrc].maxOutgoingDelay = connIt->second.maxDelay;
 
        // given group has plastic connection, and we need to apply STDP rule...
        synWtType = GET_FIXED_PLASTIC(connIt->second.connProp);
        if (synWtType == SYN_PLASTIC) {
            groupConfigMDMap[connIt->second.grpDest].fixedInputWts = false;
        }
    }
 
    // assigned global neruon ids to each group in the order...
    //    !!!!!!! IMPORTANT : NEURON ORGANIZATION/ARRANGEMENT MAP !!!!!!!!!!
    //     <--- Excitatory --> | <-------- Inhibitory REGION ----------> | <-- Excitatory -->
    //     Excitatory-Regular  | Inhibitory-Regular | Inhibitory-Poisson | Excitatory-Poisson
    int assignedGroup = 0;
    int availableNeuronId = 0;
    for(int order = 0; order < 4; order++) {
        for(int gGrpId = 0; gGrpId < numGroups; gGrpId++) {
            if (IS_EXCITATORY_TYPE(groupConfigMap[gGrpId].type) && (groupConfigMap[gGrpId].type & POISSON_NEURON) && order == 3) {
                availableNeuronId = assignGroup(gGrpId, availableNeuronId);
                assignedGroup++;
            } else if (IS_INHIBITORY_TYPE(groupConfigMap[gGrpId].type) &&  (groupConfigMap[gGrpId].type & POISSON_NEURON) && order == 2) {
                availableNeuronId = assignGroup(gGrpId, availableNeuronId);
                assignedGroup++;
            } else if (IS_EXCITATORY_TYPE(groupConfigMap[gGrpId].type) && !(groupConfigMap[gGrpId].type & POISSON_NEURON) && order == 0) {
                availableNeuronId = assignGroup(gGrpId, availableNeuronId);
                assignedGroup++;
            } else if (IS_INHIBITORY_TYPE(groupConfigMap[gGrpId].type) && !(groupConfigMap[gGrpId].type & POISSON_NEURON) && order == 1) {
                availableNeuronId = assignGroup(gGrpId, availableNeuronId);
                assignedGroup++;
            }
        }
    }
    //assert(availableNeuronId == numN);
    assert(assignedGroup == numGroups);
}
 
void SNN::connectNetwork() {
    // this parse generates local connections
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        for (std::list<ConnectConfig>::iterator connIt = localConnectLists[netId].begin(); connIt != localConnectLists[netId].end(); connIt++) {
            switch(connIt->type) {
                case CONN_RANDOM:
                    connectRandom(netId, connIt, false);
                    break;
                case CONN_FULL:
                    connectFull(netId, connIt, false);
                    break;
                case CONN_FULL_NO_DIRECT:
                    connectFull(netId, connIt, false);
                    break;
                case CONN_ONE_TO_ONE:
                    connectOneToOne(netId, connIt, false);
                    break;
                case CONN_GAUSSIAN:
                    connectGaussian(netId, connIt, false);
                    break;
                case CONN_USER_DEFINED:
                    connectUserDefined(netId, connIt, false);
                    break;
                default:
                    KERNEL_ERROR("Invalid connection type( should be 'random', 'full', 'full-no-direct', or 'one-to-one')");
                    exitSimulation(-1);
            }
        }
    }
 
    // this parse generates external connections
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        for (std::list<ConnectConfig>::iterator connIt = externalConnectLists[netId].begin(); connIt != externalConnectLists[netId].end(); connIt++) {
            switch(connIt->type) {
                case CONN_RANDOM:
                    connectRandom(netId, connIt, true);
                    break;
                case CONN_FULL:
                    connectFull(netId, connIt, true);
                    break;
                case CONN_FULL_NO_DIRECT:
                    connectFull(netId, connIt, true);
                    break;
                case CONN_ONE_TO_ONE:
                    connectOneToOne(netId, connIt, true);
                    break;
                case CONN_GAUSSIAN:
                    connectGaussian(netId, connIt, true);
                    break;
                case CONN_USER_DEFINED:
                    connectUserDefined(netId, connIt, true);
                    break;
                default:
                    KERNEL_ERROR("Invalid connection type( should be 'random', 'full', 'full-no-direct', or 'one-to-one')");
                    exitSimulation(-1);
            }
        }
    }
}
 
inline void SNN::connectNeurons(int netId, int _grpSrc, int _grpDest, int _nSrc, int _nDest, short int _connId, int externalNetId) {
    //assert(destN <= CONN_SYN_NEURON_MASK); // total number of neurons is less than 1 million within a GPU
    ConnectionInfo connInfo;
    connInfo.grpSrc = _grpSrc;
    connInfo.grpDest = _grpDest;
    connInfo.nSrc = _nSrc;
    connInfo.nDest = _nDest;
    connInfo.srcGLoffset = 0;
    connInfo.connId = _connId;
    connInfo.preSynId = -1;
    connInfo.initWt = 0.0f;
    connInfo.maxWt = 0.0f;
    connInfo.delay = 0;
 
    // generate the delay vaule
    connInfo.delay = connectConfigMap[_connId].minDelay + rand() % (connectConfigMap[_connId].maxDelay - connectConfigMap[_connId].minDelay + 1);
    assert((connInfo.delay >= connectConfigMap[_connId].minDelay) && (connInfo.delay <= connectConfigMap[_connId].maxDelay));
    // generate the max weight and initial weight
    //float initWt = generateWeight(connectConfigMap[it->connId].connProp, connectConfigMap[it->connId].initWt, connectConfigMap[it->connId].maxWt, it->nSrc, it->grpSrc);
    float initWt = connectConfigMap[_connId].initWt;
    float maxWt = connectConfigMap[_connId].maxWt;
    // adjust sign of weight based on pre-group (negative if pre is inhibitory)
    // this access is fine, isExcitatoryGroup() use global grpId
    connInfo.maxWt = isExcitatoryGroup(_grpSrc) ? fabs(maxWt) : -1.0 * fabs(maxWt);
    connInfo.initWt = isExcitatoryGroup(_grpSrc) ? fabs(initWt) : -1.0 * fabs(initWt);
 
    connectionLists[netId].push_back(connInfo);
 
    // If the connection is external, copy the connection info to the external network
    if (externalNetId >= 0)
        connectionLists[externalNetId].push_back(connInfo);
}
 
inline void SNN::connectNeurons(int netId, int _grpSrc, int _grpDest, int _nSrc, int _nDest, short int _connId, float initWt, float maxWt, uint8_t delay, int externalNetId) {
    //assert(destN <= CONN_SYN_NEURON_MASK); // total number of neurons is less than 1 million within a GPU
    ConnectionInfo connInfo;
    connInfo.grpSrc = _grpSrc;
    connInfo.grpDest = _grpDest;
    connInfo.nSrc = _nSrc;
    connInfo.nDest = _nDest;
    connInfo.srcGLoffset = 0;
    connInfo.connId = _connId;
    connInfo.preSynId = -1;
    // adjust the sign of the weight based on inh/exc connection
    connInfo.initWt = isExcitatoryGroup(_grpSrc) ? fabs(initWt) : -1.0*fabs(initWt);
    connInfo.maxWt = isExcitatoryGroup(_grpSrc) ? fabs(maxWt) : -1.0*fabs(maxWt);
    connInfo.delay = delay;
 
    connectionLists[netId].push_back(connInfo);
 
    // If the connection is external, copy the connection info to the external network
    if (externalNetId >= 0)
        connectionLists[externalNetId].push_back(connInfo);
}
 
// make 'C' full connections from grpSrc to grpDest
void SNN::connectFull(int netId, std::list<ConnectConfig>::iterator connIt, bool isExternal) {
    int grpSrc = connIt->grpSrc;
    int grpDest = connIt->grpDest;
    bool noDirect = (connIt->type == CONN_FULL_NO_DIRECT);
    int externalNetId = -1;
 
    if (isExternal) {
        externalNetId = groupConfigMDMap[grpDest].netId;
        assert(netId != externalNetId);
    }
 
    int gPreStart = groupConfigMDMap[grpSrc].gStartN;
    for(int gPreN = groupConfigMDMap[grpSrc].gStartN; gPreN <= groupConfigMDMap[grpSrc].gEndN; gPreN++)  {
        Point3D locPre = getNeuronLocation3D(grpSrc, gPreN - gPreStart); // 3D coordinates of i
        int gPostStart = groupConfigMDMap[grpDest].gStartN;
        for(int gPostN = groupConfigMDMap[grpDest].gStartN; gPostN <= groupConfigMDMap[grpDest].gEndN; gPostN++) { // j: the temp neuron id
            // if flag is set, don't connect direct connections
            if(noDirect && gPreN == gPostN)
                continue;
 
            // check whether pre-neuron location is in RF of post-neuron
            Point3D locPost = getNeuronLocation3D(grpDest, gPostN - gPostStart); // 3D coordinates of j
            if (!isPoint3DinRF(connIt->connRadius, locPre, locPost))
                continue;
 
            connectNeurons(netId, grpSrc, grpDest, gPreN, gPostN, connIt->connId, externalNetId);
            connIt->numberOfConnections++;
        }
    }
 
    std::list<GroupConfigMD>::iterator grpIt;
    GroupConfigMD targetGrp;
 
    // update numPostSynapses and numPreSynapses of groups in the local network
    targetGrp.gGrpId = grpSrc; // the other fields does not matter
    grpIt = std::find(groupPartitionLists[netId].begin(), groupPartitionLists[netId].end(), targetGrp);
    assert(grpIt != groupPartitionLists[netId].end());
    grpIt->numPostSynapses += connIt->numberOfConnections;
 
    targetGrp.gGrpId = grpDest; // the other fields does not matter
    grpIt = std::find(groupPartitionLists[netId].begin(), groupPartitionLists[netId].end(), targetGrp);
    assert(grpIt != groupPartitionLists[netId].end());
    grpIt->numPreSynapses += connIt->numberOfConnections;
 
    // also update numPostSynapses and numPreSynapses of groups in the external network if the connection is external
    if (isExternal) {
        targetGrp.gGrpId = grpSrc; // the other fields does not matter
        grpIt = std::find(groupPartitionLists[externalNetId].begin(), groupPartitionLists[externalNetId].end(), targetGrp);
        assert(grpIt != groupPartitionLists[externalNetId].end());
        grpIt->numPostSynapses += connIt->numberOfConnections;
 
        targetGrp.gGrpId = grpDest; // the other fields does not matter
        grpIt = std::find(groupPartitionLists[externalNetId].begin(), groupPartitionLists[externalNetId].end(), targetGrp);
        assert(grpIt != groupPartitionLists[externalNetId].end());
        grpIt->numPreSynapses += connIt->numberOfConnections;
    }
}
 
void SNN::connectGaussian(int netId, std::list<ConnectConfig>::iterator connIt, bool isExternal) {
    // in case pre and post have different Grid3D sizes: scale pre to the grid size of post
    int grpSrc = connIt->grpSrc;
    int grpDest = connIt->grpDest;
    Grid3D grid_i = getGroupGrid3D(grpSrc);
    Grid3D grid_j = getGroupGrid3D(grpDest);
    Point3D scalePre = Point3D(grid_j.numX, grid_j.numY, grid_j.numZ) / Point3D(grid_i.numX, grid_i.numY, grid_i.numZ);
    int externalNetId = -1;
 
    if (isExternal) {
        externalNetId = groupConfigMDMap[grpDest].netId;
        assert(netId != externalNetId);
    }
 
    for(int i = groupConfigMDMap[grpSrc].gStartN; i <= groupConfigMDMap[grpSrc].gEndN; i++)  {
        Point3D loc_i = getNeuronLocation3D(i)*scalePre; // i: adjusted 3D coordinates
 
        for(int j = groupConfigMDMap[grpDest].gStartN; j <= groupConfigMDMap[grpDest].gEndN; j++) { // j: the temp neuron id
            // check whether pre-neuron location is in RF of post-neuron
            Point3D loc_j = getNeuronLocation3D(j); // 3D coordinates of j
 
            // make sure point is in RF
            double rfDist = getRFDist3D(connIt->connRadius,loc_i,loc_j);
            if (rfDist < 0.0 || rfDist > 1.0)
                continue;
 
            // if rfDist is valid, it returns a number between 0 and 1
            // we want these numbers to fit to Gaussian weigths, so that rfDist=0 corresponds to max Gaussian weight
            // and rfDist=1 corresponds to 0.1 times max Gaussian weight
            // so we're looking at gauss = exp(-a*rfDist), where a such that exp(-a)=0.1
            // solving for a, we find that a = 2.3026
            double gauss = exp(-2.3026*rfDist);
            if (gauss < 0.1)
                continue;
 
            if (drand48() < connIt->connProbability) {
                float initWt = gauss * connIt->initWt; // scale weight according to gauss distance
                float maxWt = connIt->maxWt;
                uint8_t delay = connIt->minDelay + rand() % (connIt->maxDelay - connIt->minDelay + 1);
                assert((delay >= connIt->minDelay) && (delay <= connIt->maxDelay));
 
                connectNeurons(netId, grpSrc, grpDest, i, j, connIt->connId, initWt, maxWt, delay, externalNetId);
                connIt->numberOfConnections++;
            }
        }
    }
 
    std::list<GroupConfigMD>::iterator grpIt;
    GroupConfigMD targetGrp;
 
    // update numPostSynapses and numPreSynapses of groups in the local network
    targetGrp.gGrpId = grpSrc; // the other fields does not matter
    grpIt = std::find(groupPartitionLists[netId].begin(), groupPartitionLists[netId].end(), targetGrp);
    assert(grpIt != groupPartitionLists[netId].end());
    grpIt->numPostSynapses += connIt->numberOfConnections;
 
    targetGrp.gGrpId = grpDest; // the other fields does not matter
    grpIt = std::find(groupPartitionLists[netId].begin(), groupPartitionLists[netId].end(), targetGrp);
    assert(grpIt != groupPartitionLists[netId].end());
    grpIt->numPreSynapses += connIt->numberOfConnections;
 
    // also update numPostSynapses and numPreSynapses of groups in the external network if the connection is external
    if (isExternal) {
        targetGrp.gGrpId = grpSrc; // the other fields does not matter
        grpIt = std::find(groupPartitionLists[externalNetId].begin(), groupPartitionLists[externalNetId].end(), targetGrp);
        assert(grpIt != groupPartitionLists[externalNetId].end());
        grpIt->numPostSynapses += connIt->numberOfConnections;
 
        targetGrp.gGrpId = grpDest; // the other fields does not matter
        grpIt = std::find(groupPartitionLists[externalNetId].begin(), groupPartitionLists[externalNetId].end(), targetGrp);
        assert(grpIt != groupPartitionLists[externalNetId].end());
        grpIt->numPreSynapses += connIt->numberOfConnections;
    }
}
 
void SNN::connectOneToOne(int netId, std::list<ConnectConfig>::iterator connIt, bool isExternal) {
    int grpSrc = connIt->grpSrc;
    int grpDest = connIt->grpDest;
    int externalNetId = -1;
 
    if (isExternal) {
        externalNetId = groupConfigMDMap[grpDest].netId;
        assert(netId != externalNetId);
    }
 
    assert( groupConfigMap[grpDest].numN == groupConfigMap[grpSrc].numN);
 
    // NOTE: RadiusRF does not make a difference here: ignore
    for(int gPreN = groupConfigMDMap[grpSrc].gStartN, gPostN = groupConfigMDMap[grpDest].gStartN; gPreN <= groupConfigMDMap[grpSrc].gEndN; gPreN++, gPostN++)  {
        connectNeurons(netId, grpSrc, grpDest, gPreN, gPostN, connIt->connId, externalNetId);
        connIt->numberOfConnections++;
    }
 
    std::list<GroupConfigMD>::iterator grpIt;
    GroupConfigMD targetGrp;
 
    // update numPostSynapses and numPreSynapses of groups in the local network
    targetGrp.gGrpId = grpSrc; // the other fields does not matter
    grpIt = std::find(groupPartitionLists[netId].begin(), groupPartitionLists[netId].end(), targetGrp);
    assert(grpIt != groupPartitionLists[netId].end());
    grpIt->numPostSynapses += connIt->numberOfConnections;
 
    targetGrp.gGrpId = grpDest; // the other fields does not matter
    grpIt = std::find(groupPartitionLists[netId].begin(), groupPartitionLists[netId].end(), targetGrp);
    assert(grpIt != groupPartitionLists[netId].end());
    grpIt->numPreSynapses += connIt->numberOfConnections;
 
    // also update numPostSynapses and numPreSynapses of groups in the external network if the connection is external
    if (isExternal) {
        targetGrp.gGrpId = grpSrc; // the other fields does not matter
        grpIt = std::find(groupPartitionLists[externalNetId].begin(), groupPartitionLists[externalNetId].end(), targetGrp);
        assert(grpIt != groupPartitionLists[externalNetId].end());
        grpIt->numPostSynapses += connIt->numberOfConnections;
 
        targetGrp.gGrpId = grpDest; // the other fields does not matter
        grpIt = std::find(groupPartitionLists[externalNetId].begin(), groupPartitionLists[externalNetId].end(), targetGrp);
        assert(grpIt != groupPartitionLists[externalNetId].end());
        grpIt->numPreSynapses += connIt->numberOfConnections;
    }
}
 
// make 'C' random connections from grpSrc to grpDest
void SNN::connectRandom(int netId, std::list<ConnectConfig>::iterator connIt, bool isExternal) {
    int grpSrc = connIt->grpSrc;
    int grpDest = connIt->grpDest;
    int externalNetId = -1;
 
    if (isExternal) {
        externalNetId = groupConfigMDMap[grpDest].netId;
        assert(netId != externalNetId);
    }
 
    int gPreStart = groupConfigMDMap[grpSrc].gStartN;
    for(int gPreN = groupConfigMDMap[grpSrc].gStartN; gPreN <= groupConfigMDMap[grpSrc].gEndN; gPreN++) {
        Point3D locPre = getNeuronLocation3D(grpSrc, gPreN - gPreStart); // 3D coordinates of i
        int gPostStart = groupConfigMDMap[grpDest].gStartN;
        for(int gPostN = groupConfigMDMap[grpDest].gStartN; gPostN <= groupConfigMDMap[grpDest].gEndN; gPostN++) {
            // check whether pre-neuron location is in RF of post-neuron
            Point3D locPost = getNeuronLocation3D(grpDest, gPostN - gPostStart); // 3D coordinates of j
            if (!isPoint3DinRF(connIt->connRadius, locPre, locPost))
                continue;
 
            if (drand48() < connIt->connProbability) {
                connectNeurons(netId, grpSrc, grpDest, gPreN, gPostN, connIt->connId, externalNetId);
                connIt->numberOfConnections++;
            }
        }
    }
 
    std::list<GroupConfigMD>::iterator grpIt;
    GroupConfigMD targetGrp;
 
    // update numPostSynapses and numPreSynapses of groups in the local network
    targetGrp.gGrpId = grpSrc; // the other fields does not matter
    grpIt = std::find(groupPartitionLists[netId].begin(), groupPartitionLists[netId].end(), targetGrp);
    assert(grpIt != groupPartitionLists[netId].end());
    grpIt->numPostSynapses += connIt->numberOfConnections;
 
    targetGrp.gGrpId = grpDest; // the other fields does not matter
    grpIt = std::find(groupPartitionLists[netId].begin(), groupPartitionLists[netId].end(), targetGrp);
    assert(grpIt != groupPartitionLists[netId].end());
    grpIt->numPreSynapses += connIt->numberOfConnections;
 
    // also update numPostSynapses and numPreSynapses of groups in the external network if the connection is external
    if (isExternal) {
        targetGrp.gGrpId = grpSrc; // the other fields does not matter
        grpIt = std::find(groupPartitionLists[externalNetId].begin(), groupPartitionLists[externalNetId].end(), targetGrp);
        assert(grpIt != groupPartitionLists[externalNetId].end());
        grpIt->numPostSynapses += connIt->numberOfConnections;
 
        targetGrp.gGrpId = grpDest; // the other fields does not matter
        grpIt = std::find(groupPartitionLists[externalNetId].begin(), groupPartitionLists[externalNetId].end(), targetGrp);
        assert(grpIt != groupPartitionLists[externalNetId].end());
        grpIt->numPreSynapses += connIt->numberOfConnections;
    }
}
 
// FIXME: rewrite user-define call-back function
// user-defined functions called here...
// This is where we define our user-defined call-back function.  -- KDC
void SNN::connectUserDefined(int netId, std::list<ConnectConfig>::iterator connIt, bool isExternal) {
    int grpSrc = connIt->grpSrc;
    int grpDest = connIt->grpDest;
    int externalNetId = -1;
 
    if (isExternal) {
        externalNetId = groupConfigMDMap[grpDest].netId;
        assert(netId != externalNetId);
    }
 
    connIt->maxDelay = 0;
    int preStartN = groupConfigMDMap[grpSrc].gStartN;
    int postStartN = groupConfigMDMap[grpDest].gStartN;
    for (int pre_nid = groupConfigMDMap[grpSrc].gStartN; pre_nid <= groupConfigMDMap[grpSrc].gEndN; pre_nid++) {
        //Point3D loc_pre = getNeuronLocation3D(pre_nid); // 3D coordinates of i
        for (int post_nid = groupConfigMDMap[grpDest].gStartN; post_nid <= groupConfigMDMap[grpDest].gEndN; post_nid++) {
            float weight, maxWt, delay;
            bool connected;
 
            connIt->conn->connect(this, grpSrc, pre_nid - preStartN, grpDest, post_nid - postStartN, weight, maxWt, delay, connected);
            if (connected) {
                assert(delay >= 1);
                assert(delay <= MAX_SYN_DELAY);
                assert(abs(weight) <= abs(maxWt));
 
                if (GET_FIXED_PLASTIC(connIt->connProp) == SYN_FIXED)
                    maxWt = weight;
 
                if (fabs(maxWt) > connIt->maxWt)
                    connIt->maxWt = fabs(maxWt);
 
                if (delay > connIt->maxDelay)
                    connIt->maxDelay = delay;
 
                connectNeurons(netId, grpSrc, grpDest, pre_nid, post_nid, connIt->connId, weight, maxWt, delay, externalNetId);
                connIt->numberOfConnections++;
            }
        }
    }
 
    std::list<GroupConfigMD>::iterator grpIt;
    GroupConfigMD targetGrp;
 
    // update numPostSynapses and numPreSynapses of groups in the local network
    targetGrp.gGrpId = grpSrc; // the other fields does not matter
    grpIt = std::find(groupPartitionLists[netId].begin(), groupPartitionLists[netId].end(), targetGrp);
    assert(grpIt != groupPartitionLists[netId].end());
    grpIt->numPostSynapses += connIt->numberOfConnections;
 
    targetGrp.gGrpId = grpDest; // the other fields does not matter
    grpIt = std::find(groupPartitionLists[netId].begin(), groupPartitionLists[netId].end(), targetGrp);
    assert(grpIt != groupPartitionLists[netId].end());
    grpIt->numPreSynapses += connIt->numberOfConnections;
 
    // also update numPostSynapses and numPreSynapses of groups in the external network if the connection is external
    if (isExternal) {
        targetGrp.gGrpId = grpSrc; // the other fields does not matter
        grpIt = std::find(groupPartitionLists[externalNetId].begin(), groupPartitionLists[externalNetId].end(), targetGrp);
        assert(grpIt != groupPartitionLists[externalNetId].end());
        grpIt->numPostSynapses += connIt->numberOfConnections;
 
        targetGrp.gGrpId = grpDest; // the other fields does not matter
        grpIt = std::find(groupPartitionLists[externalNetId].begin(), groupPartitionLists[externalNetId].end(), targetGrp);
        assert(grpIt != groupPartitionLists[externalNetId].end());
        grpIt->numPreSynapses += connIt->numberOfConnections;
    }
}
 
//void SNN::connectFull(short int connId) {
//  int grpSrc = connectConfigMap[connId].grpSrc;
//  int grpDest = connectConfigMap[connId].grpDest;
//  bool noDirect = (connectConfigMap[connId].type == CONN_FULL_NO_DIRECT);
//
//  // rebuild struct for easier handling
//  RadiusRF radius(connectConfigMap[connId].radX, connectConfigMap[connId].radY, connectConfigMap[connId].radZ);
//
//  for(int i = groupConfigMap[grpSrc].StartN; i <= groupConfigMap[grpSrc].EndN; i++)  {
//      Point3D loc_i = getNeuronLocation3D(i); // 3D coordinates of i
//      for(int j = groupConfigMap[grpDest].StartN; j <= groupConfigMap[grpDest].EndN; j++) { // j: the temp neuron id
//          // if flag is set, don't connect direct connections
//          if((noDirect) && (i - groupConfigMap[grpSrc].StartN) == (j - groupConfigMap[grpDest].StartN))
//              continue;
//
//          // check whether pre-neuron location is in RF of post-neuron
//          Point3D loc_j = getNeuronLocation3D(j); // 3D coordinates of j
//          if (!isPoint3DinRF(radius, loc_i, loc_j))
//              continue;
//
//          //uint8_t dVal = info->minDelay + (int)(0.5 + (drand48() * (info->maxDelay - info->minDelay)));
//          uint8_t dVal = connectConfigMap[connId].minDelay + rand() % (connectConfigMap[connId].maxDelay - connectConfigMap[connId].minDelay + 1);
//          assert((dVal >= connectConfigMap[connId].minDelay) && (dVal <= connectConfigMap[connId].maxDelay));
//          float synWt = generateWeight(connectConfigMap[connId].connProp, connectConfigMap[connId].initWt, connectConfigMap[connId].maxWt, i, grpSrc);
//
//          setConnection(grpSrc, grpDest, i, j, synWt, connectConfigMap[connId].maxWt, dVal, connectConfigMap[connId].connProp, connId);// info->connId);
//          connectConfigMap[connId].numberOfConnections++;
//      }
//  }
//
//  groupInfo[grpSrc].sumPostConn += connectConfigMap[connId].numberOfConnections;
//  groupInfo[grpDest].sumPreConn += connectConfigMap[connId].numberOfConnections;
//}
 
//void SNN::connectGaussian(short int connId) {
//  // rebuild struct for easier handling
//  // adjust with sqrt(2) in order to make the Gaussian kernel depend on 2*sigma^2
//  RadiusRF radius(connectConfigMap[connId].radX, connectConfigMap[connId].radY, connectConfigMap[connId].radZ);
//
//  // in case pre and post have different Grid3D sizes: scale pre to the grid size of post
//  int grpSrc = connectConfigMap[connId].grpSrc;
//  int grpDest = connectConfigMap[connId].grpDest;
//  Grid3D grid_i = getGroupGrid3D(grpSrc);
//  Grid3D grid_j = getGroupGrid3D(grpDest);
//  Point3D scalePre = Point3D(grid_j.numX, grid_j.numY, grid_j.numZ) / Point3D(grid_i.numX, grid_i.numY, grid_i.numZ);
//
//  for(int i = groupConfigMap[grpSrc].StartN; i <= groupConfigMap[grpSrc].EndN; i++)  {
//      Point3D loc_i = getNeuronLocation3D(i)*scalePre; // i: adjusted 3D coordinates
//
//      for(int j = groupConfigMap[grpDest].StartN; j <= groupConfigMap[grpDest].EndN; j++) { // j: the temp neuron id
//          // check whether pre-neuron location is in RF of post-neuron
//          Point3D loc_j = getNeuronLocation3D(j); // 3D coordinates of j
//
//          // make sure point is in RF
//          double rfDist = getRFDist3D(radius,loc_i,loc_j);
//          if (rfDist < 0.0 || rfDist > 1.0)
//              continue;
//
//          // if rfDist is valid, it returns a number between 0 and 1
//          // we want these numbers to fit to Gaussian weigths, so that rfDist=0 corresponds to max Gaussian weight
//          // and rfDist=1 corresponds to 0.1 times max Gaussian weight
//          // so we're looking at gauss = exp(-a*rfDist), where a such that exp(-a)=0.1
//          // solving for a, we find that a = 2.3026
//          double gauss = exp(-2.3026*rfDist);
//          if (gauss < 0.1)
//              continue;
//
//          if (drand48() < connectConfigMap[connId].p) {
//              uint8_t dVal = connectConfigMap[connId].minDelay + rand() % (connectConfigMap[connId].maxDelay - connectConfigMap[connId].minDelay + 1);
//              assert((dVal >= connectConfigMap[connId].minDelay) && (dVal <= connectConfigMap[connId].maxDelay));
//              float synWt = gauss * connectConfigMap[connId].initWt; // scale weight according to gauss distance
//              setConnection(grpSrc, grpDest, i, j, synWt, connectConfigMap[connId].maxWt, dVal, connectConfigMap[connId].connProp, connId);//info->connId);
//              connectConfigMap[connId].numberOfConnections++;
//          }
//      }
//  }
//
//  groupInfo[grpSrc].sumPostConn += connectConfigMap[connId].numberOfConnections;
//  groupInfo[grpDest].sumPreConn += connectConfigMap[connId].numberOfConnections;
//}
//
//void SNN::connectOneToOne(short int connId) {
//  int grpSrc = connectConfigMap[connId].grpSrc;
//  int grpDest = connectConfigMap[connId].grpDest;
//  assert( groupConfigMap[grpDest].SizeN == groupConfigMap[grpSrc].SizeN );
//
//  // NOTE: RadiusRF does not make a difference here: ignore
//  for(int nid=groupConfigMap[grpSrc].StartN,j=groupConfigMap[grpDest].StartN; nid<=groupConfigMap[grpSrc].EndN; nid++, j++)  {
//      uint8_t dVal = connectConfigMap[connId].minDelay + rand() % (connectConfigMap[connId].maxDelay - connectConfigMap[connId].minDelay + 1);
//      assert((dVal >= connectConfigMap[connId].minDelay) && (dVal <= connectConfigMap[connId].maxDelay));
//      float synWt = generateWeight(connectConfigMap[connId].connProp, connectConfigMap[connId].initWt, connectConfigMap[connId].maxWt, nid, grpSrc);
//      setConnection(grpSrc, grpDest, nid, j, synWt, connectConfigMap[connId].maxWt, dVal, connectConfigMap[connId].connProp, connId);//info->connId);
//      connectConfigMap[connId].numberOfConnections++;
//  }
//
//  groupInfo[grpSrc].sumPostConn += connectConfigMap[connId].numberOfConnections;
//  groupInfo[grpDest].sumPreConn += connectConfigMap[connId].numberOfConnections;
//}
//
//void SNN::connectRandom(short int connId) {
//  int grpSrc = connectConfigMap[connId].grpSrc;
//  int grpDest = connectConfigMap[connId].grpDest;
//
//  // rebuild struct for easier handling
//  RadiusRF radius(connectConfigMap[connId].radX, connectConfigMap[connId].radY, connectConfigMap[connId].radZ);
//
//  for(int pre_nid = groupConfigMap[grpSrc].StartN; pre_nid <= groupConfigMap[grpSrc].EndN; pre_nid++) {
//      Point3D loc_pre = getNeuronLocation3D(pre_nid); // 3D coordinates of i
//      for(int post_nid = groupConfigMap[grpDest].StartN; post_nid <= groupConfigMap[grpDest].EndN; post_nid++) {
//          // check whether pre-neuron location is in RF of post-neuron
//          Point3D loc_post = getNeuronLocation3D(post_nid); // 3D coordinates of j
//          if (!isPoint3DinRF(radius, loc_pre, loc_post))
//              continue;
//
//          if (drand48() < connectConfigMap[connId].p) {
//              //uint8_t dVal = info->minDelay + (int)(0.5+(drand48()*(info->maxDelay-info->minDelay)));
//              uint8_t dVal = connectConfigMap[connId].minDelay + rand() % (connectConfigMap[connId].maxDelay - connectConfigMap[connId].minDelay + 1);
//              assert((dVal >= connectConfigMap[connId].minDelay) && (dVal <= connectConfigMap[connId].maxDelay));
//              float synWt = generateWeight(connectConfigMap[connId].connProp, connectConfigMap[connId].initWt, connectConfigMap[connId].maxWt, pre_nid, grpSrc);
//              setConnection(grpSrc, grpDest, pre_nid, post_nid, synWt, connectConfigMap[connId].maxWt, dVal, connectConfigMap[connId].connProp, connId); //info->connId);
//              connectConfigMap[connId].numberOfConnections++;
//          }
//      }
//  }
//
//  groupInfo[grpSrc].sumPostConn += connectConfigMap[connId].numberOfConnections;
//  groupInfo[grpDest].sumPreConn += connectConfigMap[connId].numberOfConnections;
//}
//
//void SNN::connectUserDefined(short int connId) {
//  int grpSrc = connectConfigMap[connId].grpSrc;
//  int grpDest = connectConfigMap[connId].grpDest;
//  connectConfigMap[connId].maxDelay = 0;
//  for(int nid=groupConfigMap[grpSrc].StartN; nid<=groupConfigMap[grpSrc].EndN; nid++) {
//      for(int nid2=groupConfigMap[grpDest].StartN; nid2 <= groupConfigMap[grpDest].EndN; nid2++) {
//          int srcId  = nid  - groupConfigMap[grpSrc].StartN;
//          int destId = nid2 - groupConfigMap[grpDest].StartN;
//          float weight, maxWt, delay;
//          bool connected;
//
//          connectConfigMap[connId].conn->connect(this, grpSrc, srcId, grpDest, destId, weight, maxWt, delay, connected);
//          if(connected)  {
//              if (GET_FIXED_PLASTIC(connectConfigMap[connId].connProp) == SYN_FIXED)
//                  maxWt = weight;
//
//              connectConfigMap[connId].maxWt = maxWt;
//
//              assert(delay >= 1);
//              assert(delay <= MAX_SYN_DELAY);
//              assert(abs(weight) <= abs(maxWt));
//
//              // adjust the sign of the weight based on inh/exc connection
//              weight = isExcitatoryGroup(grpSrc) ? fabs(weight) : -1.0*fabs(weight);
//              maxWt  = isExcitatoryGroup(grpSrc) ? fabs(maxWt)  : -1.0*fabs(maxWt);
//
//              setConnection(grpSrc, grpDest, nid, nid2, weight, maxWt, delay, connectConfigMap[connId].connProp, connId);// info->connId);
//              connectConfigMap[connId].numberOfConnections++;
//              if(delay > connectConfigMap[connId].maxDelay) {
//                  connectConfigMap[connId].maxDelay = delay;
//              }
//          }
//      }
//  }
//
//  groupInfo[grpSrc].sumPostConn += connectConfigMap[connId].numberOfConnections;
//  groupInfo[grpDest].sumPreConn += connectConfigMap[connId].numberOfConnections;
//}
 
void SNN::deleteRuntimeData() {
    // FIXME: assert simulation use GPU first
    // wait for kernels to complete
#ifndef __NO_CUDA__
    CUDA_CHECK_ERRORS(cudaThreadSynchronize());
#endif
 
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
        cpu_set_t cpus;
        ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
        int threadCount = 0;
    #endif
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            if (netId < CPU_RUNTIME_BASE) // GPU runtime
                deleteRuntimeData_GPU(netId);
            else{ // CPU runtime
                #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                    deleteRuntimeData_CPU(netId);
                #else // Linux or MAC
                    pthread_attr_t attr;
                    pthread_attr_init(&attr);
                    CPU_ZERO(&cpus);
                    CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                    pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                    argsThreadRoutine[threadCount].snn_pointer = this;
                    argsThreadRoutine[threadCount].netId = netId;
                    argsThreadRoutine[threadCount].lGrpId = 0;
                    argsThreadRoutine[threadCount].startIdx = 0;
                    argsThreadRoutine[threadCount].endIdx = 0;
                    argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                    pthread_create(&threads[threadCount], &attr, &SNN::helperDeleteRuntimeData_CPU, (void*)&argsThreadRoutine[threadCount]);
                    pthread_attr_destroy(&attr);
                    threadCount++;
                #endif
            }
        }
    }
 
    #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
        // join all the threads
        for (int i=0; i<threadCount; i++){
            pthread_join(threads[i], NULL);
        }
    #endif
 
#ifndef __NO_CUDA__
    CUDA_DELETE_TIMER(timer);
#endif
}
 
// delete all objects (CPU and GPU side)
void SNN::deleteObjects() {
    if (simulatorDeleted)
        return;
 
    printSimSummary();
 
    // deallocate objects
    resetMonitors(true);
    resetConnectionConfigs(true);
 
    // delete manager runtime data
    deleteManagerRuntimeData();
 
    deleteRuntimeData();
 
    // fclose file streams, unless in custom mode
    if (loggerMode_ != CUSTOM) {
        // don't fclose if it's stdout or stderr, otherwise they're gonna stay closed for the rest of the process
        if (fpInf_ != NULL && fpInf_ != stdout && fpInf_ != stderr)
            fclose(fpInf_);
        if (fpErr_ != NULL && fpErr_ != stdout && fpErr_ != stderr)
            fclose(fpErr_);
        if (fpDeb_ != NULL && fpDeb_ != stdout && fpDeb_ != stderr)
            fclose(fpDeb_);
        if (fpLog_ != NULL && fpLog_ != stdout && fpLog_ != stderr)
            fclose(fpLog_);
    }
 
    simulatorDeleted = true;
}
 
void SNN::findMaxNumSynapsesGroups(int* _maxNumPostSynGrp, int* _maxNumPreSynGrp) {
    *_maxNumPostSynGrp = 0;
    *_maxNumPreSynGrp = 0;
 
    //  scan all the groups and find the required information
    for (int gGrpId = 0; gGrpId < numGroups; gGrpId++) {
        // find the values for maximum postsynaptic length
        // and maximum pre-synaptic length
        if (groupConfigMDMap[gGrpId].numPostSynapses > *_maxNumPostSynGrp)
            *_maxNumPostSynGrp = groupConfigMDMap[gGrpId].numPostSynapses;
        if (groupConfigMDMap[gGrpId].numPreSynapses > *_maxNumPreSynGrp)
            *_maxNumPreSynGrp = groupConfigMDMap[gGrpId].numPreSynapses;
    }
}
 
void SNN::findMaxNumSynapsesNeurons(int _netId, int& _maxNumPostSynN, int& _maxNumPreSynN) {
    int *tempNpre, *tempNpost;
    int nSrc, nDest, numNeurons;
    std::map<int, int> globalToLocalOffset;
 
    numNeurons = networkConfigs[_netId].numNAssigned;
    tempNpre = new int[numNeurons];
    tempNpost = new int[numNeurons];
    memset(tempNpre, 0, sizeof(int) * numNeurons);
    memset(tempNpost, 0, sizeof(int) * numNeurons);
 
    // load offset between global neuron id and local neuron id
    for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[_netId].begin(); grpIt != groupPartitionLists[_netId].end(); grpIt++) {
        globalToLocalOffset[grpIt->gGrpId] = grpIt->GtoLOffset;
    }
 
    // calculate number of pre- and post- connections of each neuron
    for (std::list<ConnectionInfo>::iterator connIt = connectionLists[_netId].begin(); connIt != connectionLists[_netId].end(); connIt++) {
        nSrc = connIt->nSrc + globalToLocalOffset[connIt->grpSrc];
        nDest = connIt->nDest + globalToLocalOffset[connIt->grpDest];
        assert(nSrc < numNeurons); assert(nDest < numNeurons);
        tempNpost[nSrc]++;
        tempNpre[nDest]++;
    }
 
    // find out the maximum number of pre- and post- connections among neurons in a local network
    _maxNumPostSynN = 0;
    _maxNumPreSynN = 0;
    for (int nId = 0; nId < networkConfigs[_netId].numN; nId++) {
        if (tempNpost[nId] > _maxNumPostSynN) _maxNumPostSynN = tempNpost[nId];
        if (tempNpre[nId] > _maxNumPreSynN) _maxNumPreSynN = tempNpre[nId];
    }
 
    delete [] tempNpre;
    delete [] tempNpost;
}
 
void SNN::findMaxSpikesD1D2(int _netId, unsigned int& _maxSpikesD1, unsigned int& _maxSpikesD2) {
    _maxSpikesD1 = 0; _maxSpikesD2 = 0;
    for(std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[_netId].begin(); grpIt != groupPartitionLists[_netId].end(); grpIt++) {
        if (grpIt->maxOutgoingDelay == 1)
            _maxSpikesD1 += (groupConfigMap[grpIt->gGrpId].numN * NEURON_MAX_FIRING_RATE);
        else
            _maxSpikesD2 += (groupConfigMap[grpIt->gGrpId].numN * NEURON_MAX_FIRING_RATE);
    }
}
 
void SNN::findNumN(int _netId, int& _numN, int& _numNExternal, int& _numNAssigned,
                   int& _numNReg, int& _numNExcReg, int& _numNInhReg,
                   int& _numNPois, int& _numNExcPois, int& _numNInhPois) {
    _numN = 0; _numNExternal = 0; _numNAssigned = 0;
    _numNReg = 0; _numNExcReg = 0; _numNInhReg = 0;
    _numNPois = 0; _numNExcPois = 0; _numNInhPois = 0;
    for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[_netId].begin(); grpIt != groupPartitionLists[_netId].end(); grpIt++) {
        int sizeN = groupConfigMap[grpIt->gGrpId].numN;
        unsigned int type = groupConfigMap[grpIt->gGrpId].type;
        if (IS_EXCITATORY_TYPE(type) && (type & POISSON_NEURON) && grpIt->netId == _netId) {
            _numN += sizeN; _numNPois += sizeN; _numNExcPois += sizeN;
        } else if (IS_INHIBITORY_TYPE(type) && (type & POISSON_NEURON) && grpIt->netId == _netId) {
            _numN += sizeN; _numNPois += sizeN; _numNInhPois += sizeN;
        } else if (IS_EXCITATORY_TYPE(type) && !(type & POISSON_NEURON) && grpIt->netId == _netId) {
            _numN += sizeN; _numNReg += sizeN; _numNExcReg += sizeN;
        } else if (IS_INHIBITORY_TYPE(type) && !(type & POISSON_NEURON) && grpIt->netId == _netId) {
            _numN += sizeN; _numNReg += sizeN; _numNInhReg += sizeN;
        } else if (grpIt->netId != _netId) {
            _numNExternal += sizeN;
        } else {
            KERNEL_ERROR("Can't find catagory for the group [%d] ", grpIt->gGrpId);
            exitSimulation(-1);
        }
        _numNAssigned += sizeN;
    }
 
    assert(_numNReg == _numNExcReg + _numNInhReg);
    assert(_numNPois == _numNExcPois + _numNInhPois);
    assert(_numN == _numNReg + _numNPois);
    assert(_numNAssigned == _numN + _numNExternal);
}
 
void SNN::findNumNSpikeGenAndOffset(int _netId) {
    networkConfigs[_netId].numNSpikeGen = 0;
 
    for(int lGrpId = 0; lGrpId < networkConfigs[_netId].numGroups; lGrpId++) {
        if (_netId == groupConfigs[_netId][lGrpId].netId && groupConfigs[_netId][lGrpId].isSpikeGenerator && groupConfigs[_netId][lGrpId].isSpikeGenFunc) {
            groupConfigs[_netId][lGrpId].Noffset = networkConfigs[_netId].numNSpikeGen;
            networkConfigs[_netId].numNSpikeGen += groupConfigs[_netId][lGrpId].numN;
        }
    }
 
    assert(networkConfigs[_netId].numNSpikeGen <= networkConfigs[_netId].numNPois);
}
 
void SNN::findNumSynapsesNetwork(int _netId, int& _numPostSynNet, int& _numPreSynNet) {
    _numPostSynNet = 0;
    _numPreSynNet  = 0;
 
    for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[_netId].begin(); grpIt != groupPartitionLists[_netId].end(); grpIt++) {
        _numPostSynNet += grpIt->numPostSynapses;
        _numPreSynNet += grpIt->numPreSynapses;
        assert(_numPostSynNet < INT_MAX);
        assert(_numPreSynNet <  INT_MAX);
    }
 
    assert(_numPreSynNet == _numPostSynNet);
}
 
void SNN::fetchGroupState(int netId, int lGrpId) {
    if (netId < CPU_RUNTIME_BASE)
        copyGroupState(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], cudaMemcpyDeviceToHost, false);
    else
        copyGroupState(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], false);
}
 
void SNN::fetchWeightState(int netId, int lGrpId) {
    if (netId < CPU_RUNTIME_BASE)
        copyWeightState(netId, lGrpId, cudaMemcpyDeviceToHost);
    else
        copyWeightState(netId, lGrpId);
}
 
void SNN::fetchNeuronSpikeCount (int gGrpId) {
    if (gGrpId == ALL) {
        for (int gGrpId = 0; gGrpId < numGroups; gGrpId++) {
            fetchNeuronSpikeCount(gGrpId);
        }
    } else {
        int netId = groupConfigMDMap[gGrpId].netId;
        int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
        int LtoGOffset = groupConfigMDMap[gGrpId].LtoGOffset;
 
        if (netId < CPU_RUNTIME_BASE)
            copyNeuronSpikeCount(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], cudaMemcpyDeviceToHost, false, LtoGOffset);
        else
            copyNeuronSpikeCount(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], false, LtoGOffset);
    }
}
 
void SNN::fetchSTPState(int gGrpId) {
}
 
void SNN::fetchConductanceAMPA(int gGrpId) {
    if (gGrpId == ALL) {
        for (int gGrpId = 0; gGrpId < numGroups; gGrpId++) {
            fetchConductanceAMPA(gGrpId);
        }
    } else {
        int netId = groupConfigMDMap[gGrpId].netId;
        int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
        int LtoGOffset = groupConfigMDMap[gGrpId].LtoGOffset;
 
        if (netId < CPU_RUNTIME_BASE)
            copyConductanceAMPA(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], cudaMemcpyDeviceToHost, false, LtoGOffset);
        else
            copyConductanceAMPA(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], false, LtoGOffset);
    }
}
 
void SNN::fetchConductanceNMDA(int gGrpId) {
    if (gGrpId == ALL) {
        for (int gGrpId = 0; gGrpId < numGroups; gGrpId++) {
            fetchConductanceNMDA(gGrpId);
        }
    } else {
        int netId = groupConfigMDMap[gGrpId].netId;
        int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
        int LtoGOffset = groupConfigMDMap[gGrpId].LtoGOffset;
 
        if (netId < CPU_RUNTIME_BASE)
            copyConductanceNMDA(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], cudaMemcpyDeviceToHost, false, LtoGOffset);
        else
            copyConductanceNMDA(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], false, LtoGOffset);
    }
}
 
void SNN::fetchConductanceGABAa(int gGrpId) {
    if (gGrpId == ALL) {
        for (int gGrpId = 0; gGrpId < numGroups; gGrpId++) {
            fetchConductanceGABAa(gGrpId);
        }
    } else {
        int netId = groupConfigMDMap[gGrpId].netId;
        int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
        int LtoGOffset = groupConfigMDMap[gGrpId].LtoGOffset;
 
        if (netId < CPU_RUNTIME_BASE)
            copyConductanceGABAa(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], cudaMemcpyDeviceToHost, false, LtoGOffset);
        else
            copyConductanceGABAa(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], false, LtoGOffset);
    }
}
 
void SNN::fetchConductanceGABAb(int gGrpId) {
    if (gGrpId == ALL) {
        for (int gGrpId = 0; gGrpId < numGroups; gGrpId++) {
            fetchConductanceGABAb(gGrpId);
        }
    } else {
        int netId = groupConfigMDMap[gGrpId].netId;
        int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
        int LtoGOffset = groupConfigMDMap[gGrpId].LtoGOffset;
 
        if (netId < CPU_RUNTIME_BASE)
            copyConductanceGABAb(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], cudaMemcpyDeviceToHost, false, LtoGOffset);
        else
            copyConductanceGABAb(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], false, LtoGOffset);
    }
}
 
 
void SNN::fetchGrpIdsLookupArray(int netId) {
    if (netId < CPU_RUNTIME_BASE)
        copyGrpIdsLookupArray(netId, cudaMemcpyDeviceToHost);
    else
        copyGrpIdsLookupArray(netId);
}
 
void SNN::fetchConnIdsLookupArray(int netId) {
    if (netId < CPU_RUNTIME_BASE)
        copyConnIdsLookupArray(netId, cudaMemcpyDeviceToHost);
    else
        copyConnIdsLookupArray(netId);
}
 
void SNN::fetchLastSpikeTime(int netId) {
    if (netId < CPU_RUNTIME_BASE)
        copyLastSpikeTime(netId, cudaMemcpyDeviceToHost);
    else
        copyLastSpikeTime(netId);
}
 
void SNN::fetchPreConnectionInfo(int netId) {
    if (netId < CPU_RUNTIME_BASE)
        copyPreConnectionInfo(netId, ALL, &managerRuntimeData, &runtimeData[netId], cudaMemcpyDeviceToHost, false);
    else
        copyPreConnectionInfo(netId, ALL, &managerRuntimeData, &runtimeData[netId], false);
}
 
void SNN::fetchPostConnectionInfo(int netId) {
    if (netId < CPU_RUNTIME_BASE)
        copyPostConnectionInfo(netId, ALL, &managerRuntimeData, &runtimeData[netId], cudaMemcpyDeviceToHost, false);
    else
        copyPostConnectionInfo(netId, ALL, &managerRuntimeData, &runtimeData[netId], false);
}
 
void SNN::fetchSynapseState(int netId) {
    if (netId < CPU_RUNTIME_BASE)
        copySynapseState(netId, &managerRuntimeData, &runtimeData[netId], cudaMemcpyDeviceToHost, false);
    else
        copySynapseState(netId, &managerRuntimeData, &runtimeData[netId], false);
}
 
 
void SNN::fetchNetworkSpikeCount() {
    unsigned int spikeCountD1, spikeCountD2, spikeCountExtD1, spikeCountExtD2;
 
    managerRuntimeData.spikeCountD1 = 0;
    managerRuntimeData.spikeCountD2 = 0;
    managerRuntimeData.spikeCountExtRxD2 = 0;
    managerRuntimeData.spikeCountExtRxD1 = 0;
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
 
            if (netId < CPU_RUNTIME_BASE) {
                copyNetworkSpikeCount(netId, cudaMemcpyDeviceToHost,
                    &spikeCountD1, &spikeCountD2,
                    &spikeCountExtD1, &spikeCountExtD2);
                //printf("netId:%d, D1:%d/D2:%d, extD1:%d/D2:%d\n", netId, spikeCountD1, spikeCountD2, spikeCountExtD1, spikeCountExtD2);
            } else {
                copyNetworkSpikeCount(netId,
                    &spikeCountD1, &spikeCountD2,
                    &spikeCountExtD1, &spikeCountExtD2);
                //printf("netId:%d, D1:%d/D2:%d, extD1:%d/D2:%d\n", netId, spikeCountD1, spikeCountD2, spikeCountExtD1, spikeCountExtD2);
            }
 
            managerRuntimeData.spikeCountD2 += spikeCountD2 - spikeCountExtD2;
            managerRuntimeData.spikeCountD1 += spikeCountD1 - spikeCountExtD1;
            managerRuntimeData.spikeCountExtRxD2 += spikeCountExtD2;
            managerRuntimeData.spikeCountExtRxD1 += spikeCountExtD1;
        }
    }
 
    managerRuntimeData.spikeCount = managerRuntimeData.spikeCountD1 + managerRuntimeData.spikeCountD2;
}
 
void SNN::fetchSpikeTables(int netId) {
    if (netId < CPU_RUNTIME_BASE)
        copySpikeTables(netId, cudaMemcpyDeviceToHost);
    else
        copySpikeTables(netId);
}
 
void SNN::fetchNeuronStateBuffer(int netId, int lGrpId) {
    if (netId < CPU_RUNTIME_BASE)
        copyNeuronStateBuffer(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], cudaMemcpyDeviceToHost, false);
    else
        copyNeuronStateBuffer(netId, lGrpId, &managerRuntimeData, &runtimeData[netId], false);
}
 
void SNN::fetchExtFiringTable(int netId) {
    assert(netId < MAX_NET_PER_SNN);
 
    if (netId < CPU_RUNTIME_BASE) { // GPU runtime
        copyExtFiringTable(netId, cudaMemcpyDeviceToHost);
    } else { // CPU runtime
        copyExtFiringTable(netId);
    }
}
 
void SNN::fetchTimeTable(int netId) {
    assert(netId < MAX_NET_PER_SNN);
 
    if (netId < CPU_RUNTIME_BASE) { // GPU runtime
        copyTimeTable(netId, cudaMemcpyDeviceToHost);
    } else {
        copyTimeTable(netId, true);
    }
}
 
void SNN::writeBackTimeTable(int netId) {
    assert(netId < MAX_NET_PER_SNN);
 
    if (netId < CPU_RUNTIME_BASE) { // GPU runtime
        copyTimeTable(netId, cudaMemcpyHostToDevice);
    } else {
        copyTimeTable(netId, false);
    }
}
 
void SNN::transferSpikes(void* dest, int destNetId, void* src, int srcNetId, int size) {
#ifndef __NO_CUDA__
    if (srcNetId < CPU_RUNTIME_BASE && destNetId < CPU_RUNTIME_BASE) {
        checkAndSetGPUDevice(destNetId);
        CUDA_CHECK_ERRORS(cudaMemcpyPeer(dest, destNetId, src, srcNetId, size));
    } else if (srcNetId >= CPU_RUNTIME_BASE && destNetId < CPU_RUNTIME_BASE) {
        checkAndSetGPUDevice(destNetId);
        CUDA_CHECK_ERRORS(cudaMemcpy(dest, src, size, cudaMemcpyHostToDevice));
    } else if (srcNetId < CPU_RUNTIME_BASE && destNetId >= CPU_RUNTIME_BASE) {
        checkAndSetGPUDevice(srcNetId);
        CUDA_CHECK_ERRORS(cudaMemcpy(dest, src, size, cudaMemcpyDeviceToHost));
    } else if(srcNetId >= CPU_RUNTIME_BASE && destNetId >= CPU_RUNTIME_BASE) {
        memcpy(dest, src, size);
    }
#else
    assert(srcNetId >= CPU_RUNTIME_BASE && destNetId >= CPU_RUNTIME_BASE);
    memcpy(dest, src, size);
#endif
}
 
void SNN::convertExtSpikesD2(int netId, int startIdx, int endIdx, int GtoLOffset) {
    if (netId < CPU_RUNTIME_BASE)
        convertExtSpikesD2_GPU(netId, startIdx, endIdx, GtoLOffset);
    else
        convertExtSpikesD2_CPU(netId, startIdx, endIdx, GtoLOffset);
}
 
void SNN::convertExtSpikesD1(int netId, int startIdx, int endIdx, int GtoLOffset) {
    if (netId < CPU_RUNTIME_BASE)
        convertExtSpikesD1_GPU(netId, startIdx, endIdx, GtoLOffset);
    else
        convertExtSpikesD1_CPU(netId, startIdx, endIdx, GtoLOffset);
}
 
void SNN::routeSpikes() {
    int firingTableIdxD2, firingTableIdxD1;
    int GtoLOffset;
 
    for (std::list<RoutingTableEntry>::iterator rteItr = spikeRoutingTable.begin(); rteItr != spikeRoutingTable.end(); rteItr++) {
        int srcNetId = rteItr->srcNetId;
        int destNetId = rteItr->destNetId;
 
        fetchExtFiringTable(srcNetId);
 
        fetchTimeTable(destNetId);
        firingTableIdxD2 = managerRuntimeData.timeTableD2[simTimeMs + glbNetworkConfig.maxDelay + 1];
        firingTableIdxD1 = managerRuntimeData.timeTableD1[simTimeMs + glbNetworkConfig.maxDelay + 1];
        //KERNEL_DEBUG("GPU1 D1:%d/D2:%d", firingTableIdxD1, firingTableIdxD2);
        //printf("srcNetId %d,destNetId %d, D1:%d/D2:%d\n", srcNetId, destNetId, firingTableIdxD1, firingTableIdxD2);
 
        #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
            pthread_t threads[(2 * networkConfigs[srcNetId].numGroups) + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
            cpu_set_t cpus;
            ThreadStruct argsThreadRoutine[(2 * networkConfigs[srcNetId].numGroups) + 1]; // same as above, +1 array size
            int threadCount = 0;
        #endif
 
        for (int lGrpId = 0; lGrpId < networkConfigs[srcNetId].numGroups; lGrpId++) {
            if (groupConfigs[srcNetId][lGrpId].hasExternalConnect && managerRuntimeData.extFiringTableEndIdxD2[lGrpId] > 0) {
                // search GtoLOffset of the neural group at destination local network
                bool isFound = false;
                for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[destNetId].begin(); grpIt != groupPartitionLists[destNetId].end(); grpIt++) {
                    if (grpIt->gGrpId == groupConfigs[srcNetId][lGrpId].gGrpId) {
                        GtoLOffset = grpIt->GtoLOffset;
                        isFound = true;
                        break;
                    }
                }
 
                if (isFound) {
                    transferSpikes(runtimeData[destNetId].firingTableD2 + firingTableIdxD2, destNetId,
                        managerRuntimeData.extFiringTableD2[lGrpId], srcNetId,
                        sizeof(int) * managerRuntimeData.extFiringTableEndIdxD2[lGrpId]);
 
                    if (destNetId < CPU_RUNTIME_BASE){
                        convertExtSpikesD2_GPU(destNetId, firingTableIdxD2,
                            firingTableIdxD2 + managerRuntimeData.extFiringTableEndIdxD2[lGrpId],
                            GtoLOffset); // [StartIdx, EndIdx)
                    }
                    else{// CPU runtime
                            #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                                convertExtSpikesD2_CPU(destNetId, firingTableIdxD2,
                                    firingTableIdxD2 + managerRuntimeData.extFiringTableEndIdxD2[lGrpId],
                                    GtoLOffset); // [StartIdx, EndIdx)
                            #else // Linux or MAC
                                pthread_attr_t attr;
                                pthread_attr_init(&attr);
                                CPU_ZERO(&cpus);
                                CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                                pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                                argsThreadRoutine[threadCount].snn_pointer = this;
                                argsThreadRoutine[threadCount].netId = destNetId;
                                argsThreadRoutine[threadCount].lGrpId = 0;
                                argsThreadRoutine[threadCount].startIdx = firingTableIdxD2;
                                argsThreadRoutine[threadCount].endIdx = firingTableIdxD2 + managerRuntimeData.extFiringTableEndIdxD2[lGrpId];
                                argsThreadRoutine[threadCount].GtoLOffset = GtoLOffset;
 
                                pthread_create(&threads[threadCount], &attr, &SNN::helperConvertExtSpikesD2_CPU, (void*)&argsThreadRoutine[threadCount]);
                                pthread_attr_destroy(&attr);
                                threadCount++;
                            #endif
                    }
 
                    firingTableIdxD2 += managerRuntimeData.extFiringTableEndIdxD2[lGrpId];
                }
            }
 
            if (groupConfigs[srcNetId][lGrpId].hasExternalConnect && managerRuntimeData.extFiringTableEndIdxD1[lGrpId] > 0) {
                // search GtoLOffset of the neural group at destination local network
                bool isFound = false;
                for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[destNetId].begin(); grpIt != groupPartitionLists[destNetId].end(); grpIt++) {
                    if (grpIt->gGrpId == groupConfigs[srcNetId][lGrpId].gGrpId) {
                        GtoLOffset = grpIt->GtoLOffset;
                        isFound = true;
                        break;
                    }
                }
 
                if (isFound) {
                    transferSpikes(runtimeData[destNetId].firingTableD1 + firingTableIdxD1, destNetId,
                        managerRuntimeData.extFiringTableD1[lGrpId], srcNetId,
                        sizeof(int) * managerRuntimeData.extFiringTableEndIdxD1[lGrpId]);
                    if (destNetId < CPU_RUNTIME_BASE){
                        convertExtSpikesD1_GPU(destNetId, firingTableIdxD1,
                            firingTableIdxD1 + managerRuntimeData.extFiringTableEndIdxD1[lGrpId],
                            GtoLOffset); // [StartIdx, EndIdx)
                    }
                    else{// CPU runtime
                        #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                                convertExtSpikesD1_CPU(destNetId, firingTableIdxD1,
                                    firingTableIdxD1 + managerRuntimeData.extFiringTableEndIdxD1[lGrpId],
                                    GtoLOffset); // [StartIdx, EndIdx)
                            #else // Linux or MAC
                                pthread_attr_t attr;
                                pthread_attr_init(&attr);
                                CPU_ZERO(&cpus);
                                CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                                pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                                argsThreadRoutine[threadCount].snn_pointer = this;
                                argsThreadRoutine[threadCount].netId = destNetId;
                                argsThreadRoutine[threadCount].lGrpId = 0;
                                argsThreadRoutine[threadCount].startIdx = firingTableIdxD1;
                                argsThreadRoutine[threadCount].endIdx = firingTableIdxD1 + managerRuntimeData.extFiringTableEndIdxD1[lGrpId];
                                argsThreadRoutine[threadCount].GtoLOffset = GtoLOffset;
 
                                pthread_create(&threads[threadCount], &attr, &SNN::helperConvertExtSpikesD1_CPU, (void*)&argsThreadRoutine[threadCount]);
                                pthread_attr_destroy(&attr);
                                threadCount++;
                            #endif
                    }
                    firingTableIdxD1 += managerRuntimeData.extFiringTableEndIdxD1[lGrpId];
                }
            }
            //KERNEL_DEBUG("GPU1 New D1:%d/D2:%d", firingTableIdxD1, firingTableIdxD2);
        }
 
        #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
            // join all the threads
            for (int i=0; i<threadCount; i++){
                pthread_join(threads[i], NULL);
            }
        #endif
 
        managerRuntimeData.timeTableD2[simTimeMs + glbNetworkConfig.maxDelay + 1] = firingTableIdxD2;
        managerRuntimeData.timeTableD1[simTimeMs + glbNetworkConfig.maxDelay + 1] = firingTableIdxD1;
        writeBackTimeTable(destNetId);
    }
}
 
//We need pass the neuron id (nid) and the grpId just for the case when we want to
//ramp up/down the weights.  In that case we need to set the weights of each synapse
//depending on their nid (their position with respect to one another). -- KDC
float SNN::generateWeight(int connProp, float initWt, float maxWt, int nid, int grpId) {
    float actWts;
    //bool setRandomWeights   = GET_INITWTS_RANDOM(connProp);
    //bool setRampDownWeights = GET_INITWTS_RAMPDOWN(connProp);
    //bool setRampUpWeights   = GET_INITWTS_RAMPUP(connProp);
 
    //if (setRandomWeights)
    //  actWts = initWt * drand48();
    //else if (setRampUpWeights)
    //  actWts = (initWt + ((nid - groupConfigs[0][grpId].StartN) * (maxWt - initWt) / groupConfigs[0][grpId].SizeN));
    //else if (setRampDownWeights)
    //  actWts = (maxWt - ((nid - groupConfigs[0][grpId].StartN) * (maxWt - initWt) / groupConfigs[0][grpId].SizeN));
    //else
        actWts = initWt;
 
    return actWts;
}
 
// checks whether a connection ID contains plastic synapses O(#connections)
bool SNN::isConnectionPlastic(short int connId) {
    assert(connId != ALL);
    assert(connId < numConnections);
 
    return GET_FIXED_PLASTIC(connectConfigMap[connId].connProp);
}
 
// FIXME: distinguish the function call at CONFIG_STATE and SETUP_STATE, where groupConfigs[0][] might not be available
// or groupConfigMap is not sync with groupConfigs[0][]
// returns whether group has homeostasis enabled
bool SNN::isGroupWithHomeostasis(int grpId) {
    assert(grpId>=0 && grpId<getNumGroups());
    return (groupConfigMap[grpId].homeoConfig.WithHomeostasis);
}
 
// performs various verification checkups before building the network
void SNN::verifyNetwork() {
    // make sure number of neuron parameters have been accumulated correctly
    // NOTE: this used to be updateParameters
    //verifyNumNeurons();
 
    // make sure compartment config is valid
    verifyCompartments();
 
    // make sure STDP post-group has some incoming plastic connections
    verifySTDP();
 
    // make sure every group with homeostasis also has STDP
    verifyHomeostasis();
 
    // make sure the max delay is within bound
    assert(glbNetworkConfig.maxDelay <= MAX_SYN_DELAY);
 
    // make sure there is sufficient buffer
    //if ((networkConfigs[0].maxSpikesD1 + networkConfigs[0].maxSpikesD2) < (numNExcReg + numNInhReg + numNPois) * UNKNOWN_NEURON_MAX_FIRING_RATE) {
    //  KERNEL_ERROR("Insufficient amount of buffer allocated...");
    //  exitSimulation(1);
    //}
 
    //make sure the number of pre- and post-connection does not exceed the limitation
    //if (maxNumPostSynGrp > MAX_NUM_POST_SYN) {
    //  for (int g = 0; g < numGroups; g++) {
    //      if (groupConfigMap[g].numPostSynapses>MAX_NUM_POST_SYN)
    //          KERNEL_ERROR("Grp: %s(%d) has too many output synapses (%d), max %d.",groupInfo[g].Name.c_str(),g,
    //                      groupConfigMap[g].numPostSynapses,MAX_NUM_POST_SYN);
    //  }
    //  assert(maxNumPostSynGrp <= MAX_NUM_POST_SYN);
    //}
 
    //if (maxNumPreSynGrp > MAX_NUM_PRE_SYN) {
    //  for (int g = 0; g < numGroups; g++) {
    //      if (groupConfigMap[g].numPreSynapses>MAX_NUM_PRE_SYN)
    //          KERNEL_ERROR("Grp: %s(%d) has too many input synapses (%d), max %d.",groupInfo[g].Name.c_str(),g,
 //                         groupConfigMap[g].numPreSynapses,MAX_NUM_PRE_SYN);
    //  }
    //  assert(maxNumPreSynGrp <= MAX_NUM_PRE_SYN);
    //}
 
    // make sure maxDelay == 1 if STP is enableed
    // \FIXME: need to figure out STP buffer for delays > 1
    if (sim_with_stp && glbNetworkConfig.maxDelay > 1) {
        KERNEL_ERROR("STP with delays > 1 ms is currently not supported.");
        exitSimulation(1);
    }
 
    if (glbNetworkConfig.maxDelay > MAX_SYN_DELAY) {
        KERNEL_ERROR("You are using a synaptic delay (%d) greater than MAX_SYN_DELAY defined in config.h", glbNetworkConfig.maxDelay);
        exitSimulation(1);
    }
}
 
void SNN::verifyCompartments() {
    for (std::map<int, compConnectConfig>::iterator it = compConnectConfigMap.begin(); it != compConnectConfigMap.end(); it++)
    {
        int grpLower = it->second.grpSrc;
        int grpUpper = it->second.grpDest;
 
        // make sure groups are compartmentally enabled
        if (!groupConfigMap[grpLower].withCompartments) {
            KERNEL_ERROR("Group %s(%d) is not compartmentally enabled, cannot be part of a compartmental connection.",
                groupConfigMap[grpLower].grpName.c_str(), grpLower);
            exitSimulation(1);
        }
        if (!groupConfigMap[grpUpper].withCompartments) {
            KERNEL_ERROR("Group %s(%d) is not compartmentally enabled, cannot be part of a compartmental connection.",
                groupConfigMap[grpUpper].grpName.c_str(), grpUpper);
            exitSimulation(1);
        }
    }
}
 
// checks whether STDP is set on a post-group with incoming plastic connections
void SNN::verifySTDP() {
    for (int gGrpId=0; gGrpId<getNumGroups(); gGrpId++) {
        if (groupConfigMap[gGrpId].stdpConfig.WithSTDP) {
            // for each post-group, check if any of the incoming connections are plastic
            bool isAnyPlastic = false;
            for (std::map<int, ConnectConfig>::iterator it = connectConfigMap.begin(); it != connectConfigMap.end(); it++) {
                if (it->second.grpDest == gGrpId) {
                    // get syn wt type from connection property
                    isAnyPlastic |= GET_FIXED_PLASTIC(it->second.connProp);
                    if (isAnyPlastic) {
                        // at least one plastic connection found: break while
                        break;
                    }
                }
            }
            if (!isAnyPlastic) {
                KERNEL_ERROR("If STDP on group %d (%s) is set, group must have some incoming plastic connections.",
                    gGrpId, groupConfigMap[gGrpId].grpName.c_str());
                exitSimulation(1);
            }
        }
    }
}
 
// checks whether every group with Homeostasis also has STDP
void SNN::verifyHomeostasis() {
    for (int gGrpId=0; gGrpId<getNumGroups(); gGrpId++) {
        if (groupConfigMap[gGrpId].homeoConfig.WithHomeostasis) {
            if (!groupConfigMap[gGrpId].stdpConfig.WithSTDP) {
                KERNEL_ERROR("If homeostasis is enabled on group %d (%s), then STDP must be enabled, too.",
                    gGrpId, groupConfigMap[gGrpId].grpName.c_str());
                exitSimulation(1);
            }
        }
    }
}
 
//void SNN::verifyNumNeurons() {
//  int nExcPois = 0;
//  int nInhPois = 0;
//  int nExcReg = 0;
//  int nInhReg = 0;
//
//  //  scan all the groups and find the required information
//  //  about the group (numN, numPostSynapses, numPreSynapses and others).
//  for(int g=0; g<numGroups; g++)  {
//      if (groupConfigMap[g].Type==UNKNOWN_NEURON) {
//          KERNEL_ERROR("Unknown group for %d (%s)", g, groupInfo[g].Name.c_str());
//          exitSimulation(1);
//      }
//
//      if (IS_INHIBITORY_TYPE(groupConfigMap[g].Type) && !(groupConfigMap[g].Type & POISSON_NEURON))
//          nInhReg += groupConfigMap[g].SizeN;
//      else if (IS_EXCITATORY_TYPE(groupConfigMap[g].Type) && !(groupConfigMap[g].Type & POISSON_NEURON))
//          nExcReg += groupConfigMap[g].SizeN;
//      else if (IS_EXCITATORY_TYPE(groupConfigMap[g].Type) &&  (groupConfigMap[g].Type & POISSON_NEURON))
//          nExcPois += groupConfigMap[g].SizeN;
//      else if (IS_INHIBITORY_TYPE(groupConfigMap[g].Type) &&  (groupConfigMap[g].Type & POISSON_NEURON))
//          nInhPois += groupConfigMap[g].SizeN;
//  }
//
//  // check the newly gathered information with class members
//  if (numN != nExcReg+nInhReg+nExcPois+nInhPois) {
//      KERNEL_ERROR("nExcReg+nInhReg+nExcPois+nInhPois=%d does not add up to numN=%d",
//          nExcReg+nInhReg+nExcPois+nInhPois, numN);
//      exitSimulation(1);
//  }
//  if (numNReg != nExcReg+nInhReg) {
//      KERNEL_ERROR("nExcReg+nInhReg=%d does not add up to numNReg=%d", nExcReg+nInhReg, numNReg);
//      exitSimulation(1);
//  }
//  if (numNPois != nExcPois+nInhPois) {
//      KERNEL_ERROR("nExcPois+nInhPois=%d does not add up to numNPois=%d", nExcPois+nInhPois, numNPois);
//      exitSimulation(1);
//  }
//
//  //printf("numN=%d == %d\n",numN,nExcReg+nInhReg+nExcPois+nInhPois);
//  //printf("numNReg=%d == %d\n",numNReg, nExcReg+nInhReg);
//  //printf("numNPois=%d == %d\n",numNPois, nExcPois+nInhPois);
//
//  assert(numN <= 1000000);
//  assert((numN > 0) && (numN == numNExcReg + numNInhReg + numNPois));
//}
 
// \FIXME: not sure where this should go... maybe create some helper file?
bool SNN::isPoint3DinRF(const RadiusRF& radius, const Point3D& pre, const Point3D& post) {
    // Note: RadiusRF rad is assumed to be the fanning in to the post neuron. So if the radius is 10 pixels, it means
    // that if you look at the post neuron, it will receive input from neurons that code for locations no more than
    // 10 pixels away. (The opposite is called a response/stimulus field.)
 
    double rfDist = getRFDist3D(radius, pre, post);
    return (rfDist >= 0.0 && rfDist <= 1.0);
}
 
double SNN::getRFDist3D(const RadiusRF& radius, const Point3D& pre, const Point3D& post) {
    // Note: RadiusRF rad is assumed to be the fanning in to the post neuron. So if the radius is 10 pixels, it means
    // that if you look at the post neuron, it will receive input from neurons that code for locations no more than
    // 10 pixels away.
 
    // ready output argument
    // SNN::isPoint3DinRF() will return true (connected) if rfDist e[0.0, 1.0]
    double rfDist = -1.0;
 
    // pre and post are connected in a generic 3D ellipsoid RF if x^2/a^2 + y^2/b^2 + z^2/c^2 <= 1.0, where
    // x = pre.x-post.x, y = pre.y-post.y, z = pre.z-post.z
    // x < 0 means:  connect if y and z satisfy some constraints, but ignore x
    // x == 0 means: connect if y and z satisfy some constraints, and enforce pre.x == post.x
    if (radius.radX==0 && pre.x!=post.x || radius.radY==0 && pre.y!=post.y || radius.radZ==0 && pre.z!=post.z) {
        rfDist = -1.0;
    } else {
        // 3D ellipsoid: x^2/a^2 + y^2/b^2 + z^2/c^2 <= 1.0
        double xTerm = (radius.radX<=0) ? 0.0 : pow(pre.x-post.x,2)/pow(radius.radX,2);
        double yTerm = (radius.radY<=0) ? 0.0 : pow(pre.y-post.y,2)/pow(radius.radY,2);
        double zTerm = (radius.radZ<=0) ? 0.0 : pow(pre.z-post.z,2)/pow(radius.radZ,2);
        rfDist = xTerm + yTerm + zTerm;
    }
 
    return rfDist;
}
 
void SNN::partitionSNN() {
    int numAssignedNeurons[MAX_NET_PER_SNN] = {0};
 
    // get number of available GPU card(s) in the present machine
    numAvailableGPUs = configGPUDevice();
 
    for (std::map<int, GroupConfigMD>::iterator grpIt = groupConfigMDMap.begin(); grpIt != groupConfigMDMap.end(); grpIt++) {
        // assign a group to the GPU specified by users
        int gGrpId = grpIt->second.gGrpId;
        int netId = groupConfigMap[gGrpId].preferredNetId;
        if (netId != ANY) {
            assert(netId > ANY && netId < MAX_NET_PER_SNN);
            grpIt->second.netId = netId;
            numAssignedNeurons[netId] += groupConfigMap[gGrpId].numN;
            groupPartitionLists[netId].push_back(grpIt->second); // Copy by value, create a copy
        } else { // netId == ANY
            // TODO: add callback function that allow user to partition network by theirself
            // FIXME: make sure GPU(s) is available first
            // this parse separates groups into each local network and assign each group a netId
            if (preferredSimMode_ == CPU_MODE) {
                grpIt->second.netId = CPU_RUNTIME_BASE; // CPU 0
                numAssignedNeurons[CPU_RUNTIME_BASE] += groupConfigMap[gGrpId].numN;
                groupPartitionLists[CPU_RUNTIME_BASE].push_back(grpIt->second); // Copy by value, create a copy
            } else if (preferredSimMode_ == GPU_MODE) {
                grpIt->second.netId = GPU_RUNTIME_BASE; // GPU 0
                numAssignedNeurons[GPU_RUNTIME_BASE] += groupConfigMap[gGrpId].numN;
                groupPartitionLists[GPU_RUNTIME_BASE].push_back(grpIt->second); // Copy by value, create a copy
            } else  if (preferredSimMode_ == HYBRID_MODE) {
                // TODO: implement partition algorithm, use naive partition for now (allocate to CPU 0)
                grpIt->second.netId = CPU_RUNTIME_BASE; // CPU 0
                numAssignedNeurons[CPU_RUNTIME_BASE] += groupConfigMap[gGrpId].numN;
                groupPartitionLists[CPU_RUNTIME_BASE].push_back(grpIt->second); // Copy by value, create a copy
            } else {
                KERNEL_ERROR("Unkown simulation mode");
                exitSimulation(-1);
            }
        }
 
        if (grpIt->second.netId == -1) { // the group was not assigned to any computing backend
            KERNEL_ERROR("Can't assign the group [%d] to any partition", grpIt->second.gGrpId);
            exitSimulation(-1);
        }
    }
 
    // this parse finds local connections (i.e., connection configs that conect local groups)
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            for (std::map<int, ConnectConfig>::iterator connIt = connectConfigMap.begin(); connIt != connectConfigMap.end(); connIt++) {
                if (groupConfigMDMap[connIt->second.grpSrc].netId == netId && groupConfigMDMap[connIt->second.grpDest].netId == netId) {
                    localConnectLists[netId].push_back(connectConfigMap[connIt->second.connId]); // Copy by value
                }
            }
 
            //printf("The size of compConnectConfigMap is: %i\n", compConnectConfigMap.size());
            for (std::map<int, compConnectConfig>::iterator connIt = compConnectConfigMap.begin(); connIt != compConnectConfigMap.end(); connIt++) {
                if (groupConfigMDMap[connIt->second.grpSrc].netId == netId && groupConfigMDMap[connIt->second.grpDest].netId == netId) {
                    localCompConnectLists[netId].push_back(compConnectConfigMap[connIt->second.connId]); // Copy by value
                }
            }
        }
    }
 
    // this parse finds external groups and external connections
    spikeRoutingTable.clear();
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            for (std::map<int, ConnectConfig>::iterator connIt = connectConfigMap.begin(); connIt != connectConfigMap.end(); connIt++) {
                int srcNetId = groupConfigMDMap[connIt->second.grpSrc].netId;
                int destNetId = groupConfigMDMap[connIt->second.grpDest].netId;
                if (srcNetId == netId && destNetId != netId) {
                    // search the source group in groupPartitionLists and mark it as having external connections
                    GroupConfigMD targetGroup;
                    std::list<GroupConfigMD>::iterator srcGrpIt, destGrpIt;
 
                    targetGroup.gGrpId = connIt->second.grpSrc;
                    srcGrpIt = find(groupPartitionLists[srcNetId].begin(), groupPartitionLists[srcNetId].end(), targetGroup);
                    assert(srcGrpIt != groupPartitionLists[srcNetId].end());
                    srcGrpIt->hasExternalConnect = true;
 
                    // FIXME: fail to write external group if the only one external link across GPUs is uni directional (GPU0 -> GPU1, no GPU1 -> GPU0)
                    targetGroup.gGrpId = connIt->second.grpDest;
                    destGrpIt = find(groupPartitionLists[srcNetId].begin(), groupPartitionLists[srcNetId].end(), targetGroup);
                    if (destGrpIt == groupPartitionLists[srcNetId].end()) { // the "external" dest group has not yet been copied to te "local" group partition list
                        numAssignedNeurons[srcNetId] += groupConfigMap[connIt->second.grpDest].numN;
                        groupPartitionLists[srcNetId].push_back(groupConfigMDMap[connIt->second.grpDest]);
                    }
 
                    targetGroup.gGrpId = connIt->second.grpSrc;
                    srcGrpIt = find(groupPartitionLists[destNetId].begin(), groupPartitionLists[destNetId].end(), targetGroup);
                    if (srcGrpIt == groupPartitionLists[destNetId].end()) {
                        numAssignedNeurons[destNetId] += groupConfigMap[connIt->second.grpSrc].numN;
                        groupPartitionLists[destNetId].push_back(groupConfigMDMap[connIt->second.grpSrc]);
                    }
 
                    externalConnectLists[srcNetId].push_back(connectConfigMap[connIt->second.connId]); // Copy by value
 
                    // build the spike routing table by the way
                    //printf("%d,%d -> %d,%d\n", srcNetId, connIt->second.grpSrc, destNetId, connIt->second.grpDest);
                    RoutingTableEntry rte(srcNetId, destNetId);
                    spikeRoutingTable.push_back(rte);
                }
            }
        }
    }
 
    spikeRoutingTable.unique();
 
    // assign local neuron ids and, local group ids for each local network in the order
    // MPORTANT : NEURON ORGANIZATION/ARRANGEMENT MAP
    // <--- Excitatory --> | <-------- Inhibitory REGION ----------> | <-- Excitatory --> | <-- External -->
    // Excitatory-Regular  | Inhibitory-Regular | Inhibitory-Poisson | Excitatory-Poisson | External Neurons
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            int availableNeuronId = 0;
            int localGroupId = 0;
            for (int order = 0; order < 5; order++) {
                for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[netId].begin(); grpIt != groupPartitionLists[netId].end(); grpIt++) {
                    unsigned int type = groupConfigMap[grpIt->gGrpId].type;
                    if (IS_EXCITATORY_TYPE(type) && (type & POISSON_NEURON) && order == 3 && grpIt->netId == netId) {
                        availableNeuronId = assignGroup(grpIt, localGroupId, availableNeuronId);
                        localGroupId++;
                    } else if (IS_INHIBITORY_TYPE(type) && (type & POISSON_NEURON) && order == 2 && grpIt->netId == netId) {
                        availableNeuronId = assignGroup(grpIt, localGroupId, availableNeuronId);
                        localGroupId++;
                    } else if (IS_EXCITATORY_TYPE(type) && !(type & POISSON_NEURON) && order == 0 && grpIt->netId == netId) {
                        availableNeuronId = assignGroup(grpIt, localGroupId, availableNeuronId);
                        localGroupId++;
                    } else if (IS_INHIBITORY_TYPE(type) && !(type & POISSON_NEURON) && order == 1 && grpIt->netId == netId) {
                        availableNeuronId = assignGroup(grpIt, localGroupId, availableNeuronId);
                        localGroupId++;
                    } else if (order == 4 && grpIt->netId != netId) {
                        availableNeuronId = assignGroup(grpIt, localGroupId, availableNeuronId);
                        localGroupId++;
                    }
                }
            }
            assert(availableNeuronId == numAssignedNeurons[netId]);
            assert(localGroupId == groupPartitionLists[netId].size());
        }
    }
 
 
    // generation connections among groups according to group and connect configs
    // update ConnectConfig::numberOfConnections
    // update GroupConfig::numPostSynapses, GroupConfig::numPreSynapses
    if (loadSimFID == NULL) {
        connectNetwork();
    } else {
        KERNEL_INFO("Load Simulation");
        loadSimulation_internal(false);  // true or false doesn't matter here
    }
 
    collectGlobalNetworkConfigP();
 
    // print group and connection overview
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            KERNEL_INFO("\n+ Local Network (%d)", netId);
            KERNEL_INFO("|-+ Group List:");
            for (std::list<GroupConfigMD>::iterator grpIt = groupPartitionLists[netId].begin(); grpIt != groupPartitionLists[netId].end(); grpIt++)
                printGroupInfo(netId, grpIt);
        }
 
        if (!localConnectLists[netId].empty() || !externalConnectLists[netId].empty()) {
            KERNEL_INFO("|-+ Connection List:");
            for (std::list<ConnectConfig>::iterator connIt = localConnectLists[netId].begin(); connIt != localConnectLists[netId].end(); connIt++)
                printConnectionInfo(netId, connIt);
 
            for (std::list<ConnectConfig>::iterator connIt = externalConnectLists[netId].begin(); connIt != externalConnectLists[netId].end(); connIt++)
                printConnectionInfo(netId, connIt);
        }
    }
 
    // print spike routing table
    printSikeRoutingInfo();
 
    snnState = PARTITIONED_SNN;
}
 
int SNN::loadSimulation_internal(bool onlyPlastic) {
    // TSC: so that we can restore the file position later...
    // MB: not sure why though...
    long file_position = ftell(loadSimFID);
    
    int tmpInt;
    float tmpFloat;
 
    bool readErr = false; // keep track of reading errors
    size_t result;
 
 
    // ------- read header ----------------
 
    fseek(loadSimFID, 0, SEEK_SET);
 
    // read file signature
    result = fread(&tmpInt, sizeof(int), 1, loadSimFID);
    readErr |= (result!=1);
    if (tmpInt != 294338571) {
        KERNEL_ERROR("loadSimulation: Unknown file signature. This does not seem to be a "
            "simulation file created with CARLsim::saveSimulation.");
        exitSimulation(-1);
    }
 
    // read file version number
    result = fread(&tmpFloat, sizeof(float), 1, loadSimFID);
    readErr |= (result!=1);
    if (tmpFloat > 0.3f) {
        KERNEL_ERROR("loadSimulation: Unsupported version number (%f)",tmpFloat);
        exitSimulation(-1);
    }
 
    // read simulation time
    result = fread(&tmpFloat, sizeof(float), 1, loadSimFID);
    readErr |= (result!=1);
 
    // read execution time
    result = fread(&tmpFloat, sizeof(float), 1, loadSimFID);
    readErr |= (result!=1);
 
    // read number of neurons
    result = fread(&tmpInt, sizeof(int), 1, loadSimFID);
    readErr |= (result!=1);
    if (tmpInt != glbNetworkConfig.numN) {
        KERNEL_ERROR("loadSimulation: Number of neurons in file (%d) and simulation (%d) don't match.",
            tmpInt, glbNetworkConfig.numN);
        exitSimulation(-1);
    }
 
    // skip save and read pre-synapses & post-synapses in CARLsim5 since they are now netID based
    // read number of pre-synapses
    // result = fread(&tmpInt, sizeof(int), 1, loadSimFID);
    // readErr |= (result!=1);
    // if (numPreSynNet != tmpInt) {
    //  KERNEL_ERROR("loadSimulation: numPreSynNet in file (%d) and simulation (%d) don't match.",
    //      tmpInt, numPreSynNet);
    //  exitSimulation(-1);
    // }
 
    //result = fread(&tmpInt, sizeof(int), 1, loadSimFID);
    //readErr |= (result!=1);
    //if (numPostSynNet != tmpInt) {
    //  KERNEL_ERROR("loadSimulation: numPostSynNet in file (%d) and simulation (%d) don't match.",
    //      tmpInt, numPostSynNet);
    //  exitSimulation(-1);
    //}
 
    // read number of groups
    result = fread(&tmpInt, sizeof(int), 1, loadSimFID);
    readErr |= (result!=1);
    if (tmpInt != numGroups) {
        KERNEL_ERROR("loadSimulation: Number of groups in file (%d) and simulation (%d) don't match.",
            tmpInt, numGroups);
        exitSimulation(-1);
    }
 
    // throw reading error instead of proceeding
    if (readErr) {
        fprintf(stderr,"loadSimulation: Error while reading file header");
        exitSimulation(-1);
    }
 
 
    // ------- read group information ----------------
    for (int g=0; g<numGroups; g++) {
        // read StartN
        result = fread(&tmpInt, sizeof(int), 1, loadSimFID);
        readErr |= (result!=1);
        if (tmpInt != groupConfigMDMap[g].gStartN) {
            KERNEL_ERROR("loadSimulation: StartN in file (%d) and grpInfo (%d) for group %d don't match.",
                tmpInt, groupConfigMDMap[g].gStartN, g);
            exitSimulation(-1);
        }
 
        // read EndN
        result = fread(&tmpInt, sizeof(int), 1, loadSimFID);
        readErr |= (result!=1);
        if (tmpInt != groupConfigMDMap[g].gEndN) {
            KERNEL_ERROR("loadSimulation: EndN in file (%d) and grpInfo (%d) for group %d don't match.",
                tmpInt, groupConfigMDMap[g].gEndN, g);
            exitSimulation(-1);
        }
 
        // read SizeX
        result = fread(&tmpInt, sizeof(int), 1, loadSimFID);
        readErr |= (result!=1);
        if (tmpInt != groupConfigMap[g].grid.numX) {
            KERNEL_ERROR("loadSimulation: numX in file (%d) and grpInfo (%d) for group %d don't match.",
                tmpInt, groupConfigMap[g].grid.numX, g);
            exitSimulation(-1);
        }
 
 
        // read SizeY
        result = fread(&tmpInt, sizeof(int), 1, loadSimFID);
        readErr |= (result!=1);
        if (tmpInt != groupConfigMap[g].grid.numY) {
            KERNEL_ERROR("loadSimulation: numY in file (%d) and grpInfo (%d) for group %d don't match.",
                tmpInt, groupConfigMap[g].grid.numY, g);
            exitSimulation(-1);
        }
 
 
        // read SizeZ
        result = fread(&tmpInt, sizeof(int), 1, loadSimFID);
        readErr |= (result!=1);
        if (tmpInt != groupConfigMap[g].grid.numZ) {
            KERNEL_ERROR("loadSimulation: numZ in file (%d) and grpInfo (%d) for group %d don't match.",
                tmpInt, groupConfigMap[g].grid.numZ, g);
            exitSimulation(-1);
        }
 
 
        // read group name
        char name[100];
        result = fread(name, sizeof(char), 100, loadSimFID);
        readErr |= (result!=100);
        if (strcmp(name,groupConfigMap[g].grpName.c_str()) != 0) {
            KERNEL_ERROR("loadSimulation: Group names in file (%s) and grpInfo (%s) don't match.", name,
                groupConfigMap[g].grpName.c_str());
            exitSimulation(-1);
        }
    }
 
    if (readErr) {
        KERNEL_ERROR("loadSimulation: Error while reading group info");
        exitSimulation(-1);
    }
    //      // read weight
    //      result = fread(&weight, sizeof(float), 1, loadSimFID);
    //      readErr |= (result!=1);
 
    //      short int gIDpre = managerRuntimeData.grpIds[nIDpre];
    //      if (IS_INHIBITORY_TYPE(groupConfigs[0][gIDpre].Type) && (weight>0)
    //              || !IS_INHIBITORY_TYPE(groupConfigs[0][gIDpre].Type) && (weight<0)) {
    //          KERNEL_ERROR("loadSimulation: Sign of weight value (%s) does not match neuron type (%s)",
    //              ((weight>=0.0f)?"plus":"minus"),
    //              (IS_INHIBITORY_TYPE(groupConfigs[0][gIDpre].Type)?"inhibitory":"excitatory"));
    //          exitSimulation(-1);
    //      }
 
    //      // read max weight
    //      result = fread(&maxWeight, sizeof(float), 1, loadSimFID);
    //      readErr |= (result!=1);
    //      if (IS_INHIBITORY_TYPE(groupConfigs[0][gIDpre].Type) && (maxWeight>=0)
    //              || !IS_INHIBITORY_TYPE(groupConfigs[0][gIDpre].Type) && (maxWeight<=0)) {
    //          KERNEL_ERROR("loadSimulation: Sign of maxWeight value (%s) does not match neuron type (%s)",
    //              ((maxWeight>=0.0f)?"plus":"minus"),
    //              (IS_INHIBITORY_TYPE(groupConfigs[0][gIDpre].Type)?"inhibitory":"excitatory"));
    //          exitSimulation(-1);
    //      }
 
    // ------- read synapse information ----------------
    int net_count = 0;
    result = fread(&net_count, sizeof(int), 1, loadSimFID);
    readErr |= (result!=1);
 
    for (int i = 0; i < net_count; i++) {
        int synapse_count = 0;
        result = fread(&synapse_count, sizeof(int), 1, loadSimFID);
        for (int j = 0; j < synapse_count; j++) {
            int gGrpIdPre;
            int gGrpIdPost;
            int grpNIdPre;
            int grpNIdPost;
            int connId;
            float weight;
            float maxWeight;
            int delay;
 
            // read gGrpIdPre
            result = fread(&gGrpIdPre, sizeof(int), 1, loadSimFID);
            readErr != (result!=1);
 
            // read gGrpIdPost
            result = fread(&gGrpIdPost, sizeof(int), 1, loadSimFID);
            readErr != (result!=1);
 
            // read grpNIdPre
            result = fread(&grpNIdPre, sizeof(int), 1, loadSimFID);
            readErr != (result!=1);
 
            // read grpNIdPost
            result = fread(&grpNIdPost, sizeof(int), 1, loadSimFID);
            readErr != (result!=1);
 
            // read connId
            result = fread(&connId, sizeof(int), 1, loadSimFID);
            readErr != (result!=1);
 
            // read weight
            result = fread(&weight, sizeof(float), 1, loadSimFID);
            readErr != (result!=1);
 
            // read maxWeight
            result = fread(&maxWeight, sizeof(float), 1, loadSimFID);
            readErr != (result!=1);
 
            // read delay
            result = fread(&delay, sizeof(int), 1, loadSimFID);
            readErr != (result!=1);
 
            // check connection
            if (connectConfigMap[connId].grpSrc != gGrpIdPre) {
                KERNEL_ERROR("loadSimulation: source group in file (%d) and in simulation (%d) for connection %d don't match.",
                    gGrpIdPre , connectConfigMap[connId].grpSrc, connId);
                exitSimulation(-1);
            }
 
            if (connectConfigMap[connId].grpDest != gGrpIdPost) {
                KERNEL_ERROR("loadSimulation: dest group in file (%d) and in simulation (%d) for connection %d don't match.",
                    gGrpIdPost , connectConfigMap[connId].grpDest, connId);
                exitSimulation(-1);
            }
 
            // connect synapse
            // find netid for two groups
            int netIdPre = groupConfigMDMap[gGrpIdPre].netId;
            int netIdPost = groupConfigMDMap[gGrpIdPost].netId;
            bool isExternal = (netIdPre != netIdPost);
 
            // find global neuron id for two neurons
            int globalNIdPre = groupConfigMDMap[gGrpIdPre].gStartN + grpNIdPre;
            int globalNIdPost = groupConfigMDMap[gGrpIdPost].gStartN + grpNIdPost;
 
            bool connected =false;
            if (!isExternal) {
                for (std::list<ConnectConfig>::iterator connIt = localConnectLists[netIdPre].begin(); connIt != localConnectLists[netIdPre].end() && (!connected); connIt++) {
                    if (connIt->connId == connId) {
                        // connect two neurons
                        connectNeurons(netIdPre, gGrpIdPre, gGrpIdPost, globalNIdPre, globalNIdPost, connId, weight, maxWeight, delay,  -1);
                        connected = true;
                        // update connection information
                        connIt->numberOfConnections++;
                        std::list<GroupConfigMD>::iterator grpIt;
 
                        // fix me maybe: numPostSynapses and numPreSynpases could also be loaded from saved information directly to save time
                        // the current implementation is a safer one
                        GroupConfigMD targetGrp;
 
                        targetGrp.gGrpId = gGrpIdPre;
                        grpIt = std::find(groupPartitionLists[netIdPre].begin(), groupPartitionLists[netIdPre].end(), targetGrp);
                        assert(grpIt != groupPartitionLists[netIdPre].end());
                        grpIt->numPostSynapses += 1;
 
                        targetGrp.gGrpId = gGrpIdPost;
                        grpIt = std::find(groupPartitionLists[netIdPre].begin(), groupPartitionLists[netIdPre].end(), targetGrp);
                        assert(grpIt != groupPartitionLists[netIdPost].end());
                        grpIt->numPreSynapses += 1;
                    }
                }
            } else {
                for (std::list<ConnectConfig>::iterator connIt = externalConnectLists[netIdPre].begin(); connIt != externalConnectLists[netIdPre].end() && (!connected); connIt++) {
                    if (connIt->connId == connId) {
                        // connect two neurons
                        connectNeurons(netIdPre, gGrpIdPre, gGrpIdPost, globalNIdPre, globalNIdPost, connId, weight, maxWeight, delay, netIdPost);
                        connected = true;
                        // update connection information
                        connIt->numberOfConnections++;
 
                        // fix me maybe: numPostSynapses and numPreSynpases could also be loaded from saved information directly to save time
                        // the current implementation is a safer one
                        GroupConfigMD targetGrp;
                        std::list<GroupConfigMD>::iterator grpIt;
 
                        targetGrp.gGrpId = gGrpIdPre;
                        grpIt = std::find(groupPartitionLists[netIdPre].begin(), groupPartitionLists[netIdPre].end(), targetGrp);
                        assert(grpIt != groupPartitionLists[netIdPre].end());
                        grpIt->numPostSynapses += 1;
 
                        targetGrp.gGrpId = gGrpIdPost;
                        grpIt = std::find(groupPartitionLists[netIdPre].begin(), groupPartitionLists[netIdPre].end(), targetGrp);
                        assert(grpIt != groupPartitionLists[netIdPost].end());
                        grpIt->numPreSynapses += 1;
 
                        // update group information in another network
                        targetGrp.gGrpId = gGrpIdPre;
                        grpIt = std::find(groupPartitionLists[netIdPost].begin(), groupPartitionLists[netIdPost].end(), targetGrp);
                        assert(grpIt != groupPartitionLists[netIdPost].end());
                        grpIt->numPostSynapses += 1;
 
                        targetGrp.gGrpId = gGrpIdPost;
                        grpIt = std::find(groupPartitionLists[netIdPost].begin(), groupPartitionLists[netIdPost].end(), targetGrp);
                        assert(grpIt != groupPartitionLists[netIdPost].end());
                        grpIt->numPreSynapses += 1;
                    }
                }
            }
        }
    }
 
    fseek(loadSimFID,file_position,SEEK_SET);
 
    return 0;
}
 
void SNN::generateRuntimeSNN() {
    // 1. genearte configurations for the simulation
    // generate (copy) group configs from groupPartitionLists[]
    generateRuntimeGroupConfigs();
 
    // generate (copy) connection configs from localConnectLists[] and exeternalConnectLists[]
    generateRuntimeConnectConfigs();
 
    // generate local network configs and accquire maximum size of rumtime data
    generateRuntimeNetworkConfigs();
 
    // 2. allocate space of runtime data used by the manager
    // - allocate firingTableD1, firingTableD2, timeTableD1, timeTableD2
    // - reset firingTableD1, firingTableD2, timeTableD1, timeTableD2
    allocateManagerSpikeTables();
    // - allocate voltage, recovery, Izh_a, Izh_b, Izh_c, Izh_d, current, extCurrent, gAMPA, gNMDA, gGABAa, gGABAb
    // lastSpikeTime, nSpikeCnt, stpu, stpx, Npre, Npre_plastic, Npost, cumulativePost, cumulativePre,
    // postSynapticIds, postDelayInfo, wt, wtChange, synSpikeTime, maxSynWt, preSynapticIds, grpIds, connIdsPreIdx,
    // grpDA, grp5HT, grpACh, grpNE, grpDABuffer, grp5HTBuffer, grpAChBuffer, grpNEBuffer, mulSynFast, mulSynSlow
    // - reset all above
    allocateManagerRuntimeData();
 
    // 3. initialize manager runtime data according to partitions (i.e., local networks)
    // 4a. allocate appropriate memory space (e.g., main memory (CPU) or device memory (GPU)).
    // 4b. load (copy) them to appropriate memory space for execution
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            KERNEL_INFO("");
            if (netId < CPU_RUNTIME_BASE) {
                KERNEL_INFO("*****************      Initializing GPU %d Runtime      *************************", netId);
            } else {
                KERNEL_INFO("*****************      Initializing CPU %d Runtime      *************************", (netId - CPU_RUNTIME_BASE));
            }
            // build the runtime data according to local network, group, connection configuirations
 
            // generate runtime data for each group
            for(int lGrpId = 0; lGrpId < networkConfigs[netId].numGroups; lGrpId++) {
                // local poisson neurons
                if (groupConfigs[netId][lGrpId].netId == netId && (groupConfigs[netId][lGrpId].Type & POISSON_NEURON)) {
                    // - init lstSpikeTime
                    // - reset avgFiring, stpu, stpx
                    // - init stpx
                    generatePoissonGroupRuntime(netId, lGrpId);
                }
                // local regular neurons
                if (groupConfigs[netId][lGrpId].netId == netId && !(groupConfigs[netId][lGrpId].Type & POISSON_NEURON)) {
                    // - init grpDA, grp5HT, grpACh, grpNE
                    // - init Izh_a, Izh_b, Izh_c, Izh_d, voltage, recovery, stpu, stpx
                    // - init baseFiring, avgFiring
                    // - init lastSpikeTime
                    generateGroupRuntime(netId, lGrpId);
                }
            }
 
            // - init grpIds
            for (int lNId = 0; lNId < networkConfigs[netId].numNAssigned; lNId++) {
                managerRuntimeData.grpIds[lNId] = -1;
                for(int lGrpId = 0; lGrpId < networkConfigs[netId].numGroupsAssigned; lGrpId++) {
                    if (lNId >= groupConfigs[netId][lGrpId].lStartN && lNId <= groupConfigs[netId][lGrpId].lEndN) {
                        managerRuntimeData.grpIds[lNId] = (short int)lGrpId;
                        break;
                    }
                }
                assert(managerRuntimeData.grpIds[lNId] != -1);
            }
 
            // - init mulSynFast, mulSynSlow
            // - init Npre, Npre_plastic, Npost, cumulativePre, cumulativePost, preSynapticIds, postSynapticIds, postDelayInfo
            // - init wt, maxSynWt
            generateConnectionRuntime(netId);
 
            generateCompConnectionRuntime(netId);
 
            // - reset current
            resetCurrent(netId);
            // - reset conductance
            resetConductances(netId);
 
            // - reset wtChange
            // - init synSpikeTime
            resetSynapse(netId, false);
 
            allocateSNN(netId);
        }
    }
 
    // count allocated CPU/GPU runtime
    numGPUs = 0; numCores = 0;
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (netId < CPU_RUNTIME_BASE && runtimeData[netId].allocated)
            numGPUs++;
        if (netId >= CPU_RUNTIME_BASE && runtimeData[netId].allocated)
            numCores++;
    }
 
    // 5. declare the spiking neural network is excutable
    snnState = EXECUTABLE_SNN;
}
 
void SNN::resetConductances(int netId) {
    if (networkConfigs[netId].sim_with_conductances) {
        memset(managerRuntimeData.gAMPA, 0, sizeof(float) * networkConfigs[netId].numNReg);
        if (networkConfigs[netId].sim_with_NMDA_rise) {
            memset(managerRuntimeData.gNMDA_r, 0, sizeof(float) * networkConfigs[netId].numNReg);
            memset(managerRuntimeData.gNMDA_d, 0, sizeof(float) * networkConfigs[netId].numNReg);
        } else {
            memset(managerRuntimeData.gNMDA, 0, sizeof(float) * networkConfigs[netId].numNReg);
        }
        memset(managerRuntimeData.gGABAa, 0, sizeof(float) * networkConfigs[netId].numNReg);
        if (networkConfigs[netId].sim_with_GABAb_rise) {
            memset(managerRuntimeData.gGABAb_r, 0, sizeof(float) * networkConfigs[netId].numNReg);
            memset(managerRuntimeData.gGABAb_d, 0, sizeof(float) * networkConfigs[netId].numNReg);
        } else {
            memset(managerRuntimeData.gGABAb, 0, sizeof(float) * networkConfigs[netId].numNReg);
        }
    }
}
 
void SNN::resetCurrent(int netId) {
    assert(managerRuntimeData.current != NULL);
    memset(managerRuntimeData.current, 0, sizeof(float) * networkConfigs[netId].numNReg);
}
 
// FIXME: unused function
void SNN::resetFiringInformation() {
    // Reset firing tables and time tables to default values..
 
    // reset various times...
    simTimeMs  = 0;
    simTimeSec = 0;
    simTime    = 0;
 
    // reset the propogation Buffer.
    resetPropogationBuffer();
    // reset Timing  Table..
    resetTimeTable();
}
 
void SNN::resetTiming() {
    prevExecutionTime = cumExecutionTime;
    executionTime = 0.0f;
}
 
void SNN::resetNeuromodulator(int netId, int lGrpId) {
    managerRuntimeData.grpDA[lGrpId] = groupConfigs[netId][lGrpId].baseDP;
    managerRuntimeData.grp5HT[lGrpId] = groupConfigs[netId][lGrpId].base5HT;
    managerRuntimeData.grpACh[lGrpId] = groupConfigs[netId][lGrpId].baseACh;
    managerRuntimeData.grpNE[lGrpId] = groupConfigs[netId][lGrpId].baseNE;
}
 
void SNN::resetNeuron(int netId, int lGrpId, int lNId) {
    int gGrpId = groupConfigs[netId][lGrpId].gGrpId; // get global group id
    assert(lNId < networkConfigs[netId].numNReg);
 
    if (groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_a == -1 && groupConfigMap[gGrpId].isLIF == 0) {
        KERNEL_ERROR("setNeuronParameters must be called for group %s (G:%d,L:%d)",groupConfigMap[gGrpId].grpName.c_str(), gGrpId, lGrpId);
        exitSimulation(1);
    }
 
    if (groupConfigMap[gGrpId].neuralDynamicsConfig.lif_tau_m == -1 && groupConfigMap[gGrpId].isLIF == 1) {
        KERNEL_ERROR("setNeuronParametersLIF must be called for group %s (G:%d,L:%d)",groupConfigMap[gGrpId].grpName.c_str(), gGrpId, lGrpId);
        exitSimulation(1);
    }
 
    managerRuntimeData.Izh_a[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_a + groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_a_sd * (float)drand48();
    managerRuntimeData.Izh_b[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_b + groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_b_sd * (float)drand48();
    managerRuntimeData.Izh_c[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_c + groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_c_sd * (float)drand48();
    managerRuntimeData.Izh_d[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_d + groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_d_sd * (float)drand48();
    managerRuntimeData.Izh_C[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_C + groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_C_sd * (float)drand48();
    managerRuntimeData.Izh_k[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_k + groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_k_sd * (float)drand48();
    managerRuntimeData.Izh_vr[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vr + groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vr_sd * (float)drand48();
    managerRuntimeData.Izh_vt[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vt + groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vt_sd * (float)drand48();
    managerRuntimeData.Izh_vpeak[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vpeak + groupConfigMap[gGrpId].neuralDynamicsConfig.Izh_vpeak_sd * (float)drand48();
    managerRuntimeData.lif_tau_m[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.lif_tau_m;
    managerRuntimeData.lif_tau_ref[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.lif_tau_ref;
    managerRuntimeData.lif_tau_ref_c[lNId] = 0;
    managerRuntimeData.lif_vTh[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.lif_vTh;
    managerRuntimeData.lif_vReset[lNId] = groupConfigMap[gGrpId].neuralDynamicsConfig.lif_vReset;
 
    // calculate gain and bias for the lif neuron
    if (groupConfigs[netId][lGrpId].isLIF){
        // gain an bias of the LIF neuron is calculated based on Membrane resistance
        float rmRange = (float)(groupConfigMap[gGrpId].neuralDynamicsConfig.lif_maxRmem - groupConfigMap[gGrpId].neuralDynamicsConfig.lif_minRmem);
        float minRmem = (float)groupConfigMap[gGrpId].neuralDynamicsConfig.lif_minRmem;
        managerRuntimeData.lif_bias[lNId] = 0.0f;
        managerRuntimeData.lif_gain[lNId] = minRmem + rmRange * (float)drand48();
    }
 
    managerRuntimeData.nextVoltage[lNId] = managerRuntimeData.voltage[lNId] = groupConfigs[netId][lGrpId].isLIF ? managerRuntimeData.lif_vReset[lNId] : (groupConfigs[netId][lGrpId].withParamModel_9 ? managerRuntimeData.Izh_vr[lNId] : managerRuntimeData.Izh_c[lNId]);
    managerRuntimeData.recovery[lNId] = groupConfigs[netId][lGrpId].withParamModel_9 ? 0.0f : managerRuntimeData.Izh_b[lNId] * managerRuntimeData.voltage[lNId];
 
    if (groupConfigs[netId][lGrpId].WithHomeostasis) {
        // set the baseFiring with some standard deviation.
        if (drand48() > 0.5) {
            managerRuntimeData.baseFiring[lNId] = groupConfigMap[gGrpId].homeoConfig.baseFiring + groupConfigMap[gGrpId].homeoConfig.baseFiringSD * -log(drand48());
        } else {
            managerRuntimeData.baseFiring[lNId] = groupConfigMap[gGrpId].homeoConfig.baseFiring - groupConfigMap[gGrpId].homeoConfig.baseFiringSD * -log(drand48());
            if(managerRuntimeData.baseFiring[lNId] < 0.1f) managerRuntimeData.baseFiring[lNId] = 0.1f;
        }
 
        if (groupConfigMap[gGrpId].homeoConfig.baseFiring != 0.0f) {
            managerRuntimeData.avgFiring[lNId] = managerRuntimeData.baseFiring[lNId];
        } else {
            managerRuntimeData.baseFiring[lNId] = 0.0f;
            managerRuntimeData.avgFiring[lNId]  = 0.0f;
        }
    }
 
    managerRuntimeData.lastSpikeTime[lNId] = MAX_SIMULATION_TIME;
 
    if(groupConfigs[netId][lGrpId].WithSTP) {
        for (int j = 0; j < networkConfigs[netId].maxDelay + 1; j++) { // is of size maxDelay_+1
            int index = STP_BUF_POS(lNId, j, networkConfigs[netId].maxDelay);
            managerRuntimeData.stpu[index] = 0.0f;
            managerRuntimeData.stpx[index] = 1.0f;
        }
    }
}
 
void SNN::resetMonitors(bool deallocate) {
    // order is important! monitor objects might point to SNN or CARLsim,
    // need to deallocate them first
 
 
    // -------------- DEALLOCATE MONITOR OBJECTS ---------------------- //
 
    // delete all SpikeMonitor objects
    // don't kill SpikeMonitorCore objects, they will get killed automatically
    for (int i=0; i<numSpikeMonitor; i++) {
        if (spikeMonList[i]!=NULL && deallocate) delete spikeMonList[i];
        spikeMonList[i]=NULL;
    }
 
    // delete all NeuronMonitor objects
    // don't kill NeuronMonitorCore objects, they will get killed automatically
    for (int i = 0; i<numNeuronMonitor; i++) {
        if (neuronMonList[i] != NULL && deallocate) delete neuronMonList[i];
        neuronMonList[i] = NULL;
    }
 
    // delete all GroupMonitor objects
    // don't kill GroupMonitorCore objects, they will get killed automatically
    for (int i=0; i<numGroupMonitor; i++) {
        if (groupMonList[i]!=NULL && deallocate) delete groupMonList[i];
        groupMonList[i]=NULL;
    }
 
    // delete all ConnectionMonitor objects
    // don't kill ConnectionMonitorCore objects, they will get killed automatically
    for (int i=0; i<numConnectionMonitor; i++) {
        if (connMonList[i]!=NULL && deallocate) delete connMonList[i];
        connMonList[i]=NULL;
    }
}
 
void SNN::resetGroupConfigs(bool deallocate) {
    // clear all existing group configurations
    if (deallocate) groupConfigMap.clear();
}
 
void SNN::resetConnectionConfigs(bool deallocate) {
    // clear all existing connection configurations
    if (deallocate) connectConfigMap.clear();
}
 
void SNN::deleteManagerRuntimeData() {
    if (spikeBuf!=NULL) delete spikeBuf;
    if (managerRuntimeData.spikeGenBits!=NULL) delete[] managerRuntimeData.spikeGenBits;
    spikeBuf=NULL; managerRuntimeData.spikeGenBits=NULL;
 
    // clear data (i.e., concentration of neuromodulator) of groups
    if (managerRuntimeData.grpDA != NULL) delete [] managerRuntimeData.grpDA;
    if (managerRuntimeData.grp5HT != NULL) delete [] managerRuntimeData.grp5HT;
    if (managerRuntimeData.grpACh != NULL) delete [] managerRuntimeData.grpACh;
    if (managerRuntimeData.grpNE != NULL) delete [] managerRuntimeData.grpNE;
    managerRuntimeData.grpDA = NULL;
    managerRuntimeData.grp5HT = NULL;
    managerRuntimeData.grpACh = NULL;
    managerRuntimeData.grpNE = NULL;
 
    // clear assistive data buffer for group monitor
    if (managerRuntimeData.grpDABuffer != NULL) delete [] managerRuntimeData.grpDABuffer;
    if (managerRuntimeData.grp5HTBuffer != NULL) delete [] managerRuntimeData.grp5HTBuffer;
    if (managerRuntimeData.grpAChBuffer != NULL) delete [] managerRuntimeData.grpAChBuffer;
    if (managerRuntimeData.grpNEBuffer != NULL) delete [] managerRuntimeData.grpNEBuffer;
    managerRuntimeData.grpDABuffer = NULL; managerRuntimeData.grp5HTBuffer = NULL;
    managerRuntimeData.grpAChBuffer = NULL; managerRuntimeData.grpNEBuffer = NULL;
 
    // -------------- DEALLOCATE CORE OBJECTS ---------------------- //
 
    if (managerRuntimeData.voltage!=NULL) delete[] managerRuntimeData.voltage;
    if (managerRuntimeData.nextVoltage != NULL) delete[] managerRuntimeData.nextVoltage;
    if (managerRuntimeData.recovery!=NULL) delete[] managerRuntimeData.recovery;
    if (managerRuntimeData.current!=NULL) delete[] managerRuntimeData.current;
    if (managerRuntimeData.extCurrent!=NULL) delete[] managerRuntimeData.extCurrent;
    if (managerRuntimeData.totalCurrent != NULL) delete[] managerRuntimeData.totalCurrent;
    if (managerRuntimeData.curSpike != NULL) delete[] managerRuntimeData.curSpike;
    if (managerRuntimeData.nVBuffer != NULL) delete[] managerRuntimeData.nVBuffer;
    if (managerRuntimeData.nUBuffer != NULL) delete[] managerRuntimeData.nUBuffer;
    if (managerRuntimeData.nIBuffer != NULL) delete[] managerRuntimeData.nIBuffer;
    managerRuntimeData.voltage=NULL; managerRuntimeData.recovery=NULL; managerRuntimeData.current=NULL; managerRuntimeData.extCurrent=NULL;
    managerRuntimeData.nextVoltage = NULL; managerRuntimeData.totalCurrent = NULL; managerRuntimeData.curSpike = NULL;
    managerRuntimeData.nVBuffer = NULL; managerRuntimeData.nUBuffer = NULL; managerRuntimeData.nIBuffer = NULL;
 
    if (managerRuntimeData.Izh_a!=NULL) delete[] managerRuntimeData.Izh_a;
    if (managerRuntimeData.Izh_b!=NULL) delete[] managerRuntimeData.Izh_b;
    if (managerRuntimeData.Izh_c!=NULL) delete[] managerRuntimeData.Izh_c;
    if (managerRuntimeData.Izh_d!=NULL) delete[] managerRuntimeData.Izh_d;
    if (managerRuntimeData.Izh_C!=NULL) delete[] managerRuntimeData.Izh_C;
    if (managerRuntimeData.Izh_k!=NULL) delete[] managerRuntimeData.Izh_k;
    if (managerRuntimeData.Izh_vr!=NULL) delete[] managerRuntimeData.Izh_vr;
    if (managerRuntimeData.Izh_vt!=NULL) delete[] managerRuntimeData.Izh_vt;
    if (managerRuntimeData.Izh_vpeak!=NULL) delete[] managerRuntimeData.Izh_vpeak;
    managerRuntimeData.Izh_a=NULL; managerRuntimeData.Izh_b=NULL; managerRuntimeData.Izh_c=NULL; managerRuntimeData.Izh_d=NULL;
    managerRuntimeData.Izh_C = NULL; managerRuntimeData.Izh_k = NULL; managerRuntimeData.Izh_vr = NULL; managerRuntimeData.Izh_vt = NULL; managerRuntimeData.Izh_vpeak = NULL;
 
    if (managerRuntimeData.lif_tau_m!=NULL) delete[] managerRuntimeData.lif_tau_m;
    if (managerRuntimeData.lif_tau_ref!=NULL) delete[] managerRuntimeData.lif_tau_ref;
    if (managerRuntimeData.lif_tau_ref_c!=NULL) delete[] managerRuntimeData.lif_tau_ref_c;
    if (managerRuntimeData.lif_vTh!=NULL) delete[] managerRuntimeData.lif_vTh;
    if (managerRuntimeData.lif_vReset!=NULL) delete[] managerRuntimeData.lif_vReset;
    if (managerRuntimeData.lif_gain!=NULL) delete[] managerRuntimeData.lif_gain;
    if (managerRuntimeData.lif_bias!=NULL) delete[] managerRuntimeData.lif_bias;
    managerRuntimeData.lif_tau_m=NULL; managerRuntimeData.lif_tau_ref=NULL; managerRuntimeData.lif_vTh=NULL;
    managerRuntimeData.lif_vReset=NULL; managerRuntimeData.lif_gain=NULL; managerRuntimeData.lif_bias=NULL;
    managerRuntimeData.lif_tau_ref_c=NULL;
 
    if (managerRuntimeData.Npre!=NULL) delete[] managerRuntimeData.Npre;
    if (managerRuntimeData.Npre_plastic!=NULL) delete[] managerRuntimeData.Npre_plastic;
    if (managerRuntimeData.Npost!=NULL) delete[] managerRuntimeData.Npost;
    managerRuntimeData.Npre=NULL; managerRuntimeData.Npre_plastic=NULL; managerRuntimeData.Npost=NULL;
 
    if (managerRuntimeData.cumulativePre!=NULL) delete[] managerRuntimeData.cumulativePre;
    if (managerRuntimeData.cumulativePost!=NULL) delete[] managerRuntimeData.cumulativePost;
    managerRuntimeData.cumulativePre=NULL; managerRuntimeData.cumulativePost=NULL;
 
    if (managerRuntimeData.gAMPA!=NULL) delete[] managerRuntimeData.gAMPA;
    if (managerRuntimeData.gNMDA!=NULL) delete[] managerRuntimeData.gNMDA;
    if (managerRuntimeData.gNMDA_r!=NULL) delete[] managerRuntimeData.gNMDA_r;
    if (managerRuntimeData.gNMDA_d!=NULL) delete[] managerRuntimeData.gNMDA_d;
    if (managerRuntimeData.gGABAa!=NULL) delete[] managerRuntimeData.gGABAa;
    if (managerRuntimeData.gGABAb!=NULL) delete[] managerRuntimeData.gGABAb;
    if (managerRuntimeData.gGABAb_r!=NULL) delete[] managerRuntimeData.gGABAb_r;
    if (managerRuntimeData.gGABAb_d!=NULL) delete[] managerRuntimeData.gGABAb_d;
    managerRuntimeData.gAMPA=NULL; managerRuntimeData.gNMDA=NULL; managerRuntimeData.gNMDA_r=NULL; managerRuntimeData.gNMDA_d=NULL;
    managerRuntimeData.gGABAa=NULL; managerRuntimeData.gGABAb=NULL; managerRuntimeData.gGABAb_r=NULL; managerRuntimeData.gGABAb_d=NULL;
 
    if (managerRuntimeData.stpu!=NULL) delete[] managerRuntimeData.stpu;
    if (managerRuntimeData.stpx!=NULL) delete[] managerRuntimeData.stpx;
    managerRuntimeData.stpu=NULL; managerRuntimeData.stpx=NULL;
 
    if (managerRuntimeData.avgFiring!=NULL) delete[] managerRuntimeData.avgFiring;
    if (managerRuntimeData.baseFiring!=NULL) delete[] managerRuntimeData.baseFiring;
    managerRuntimeData.avgFiring=NULL; managerRuntimeData.baseFiring=NULL;
 
    if (managerRuntimeData.lastSpikeTime!=NULL) delete[] managerRuntimeData.lastSpikeTime;
    if (managerRuntimeData.synSpikeTime !=NULL) delete[] managerRuntimeData.synSpikeTime;
    if (managerRuntimeData.nSpikeCnt!=NULL) delete[] managerRuntimeData.nSpikeCnt;
    managerRuntimeData.lastSpikeTime=NULL; managerRuntimeData.synSpikeTime=NULL; managerRuntimeData.nSpikeCnt=NULL;
 
    if (managerRuntimeData.postDelayInfo!=NULL) delete[] managerRuntimeData.postDelayInfo;
    if (managerRuntimeData.preSynapticIds!=NULL) delete[] managerRuntimeData.preSynapticIds;
    if (managerRuntimeData.postSynapticIds!=NULL) delete[] managerRuntimeData.postSynapticIds;
    managerRuntimeData.postDelayInfo=NULL; managerRuntimeData.preSynapticIds=NULL; managerRuntimeData.postSynapticIds=NULL;
 
    if (managerRuntimeData.wt!=NULL) delete[] managerRuntimeData.wt;
    if (managerRuntimeData.maxSynWt!=NULL) delete[] managerRuntimeData.maxSynWt;
    if (managerRuntimeData.wtChange !=NULL) delete[] managerRuntimeData.wtChange;
    managerRuntimeData.wt=NULL; managerRuntimeData.maxSynWt=NULL; managerRuntimeData.wtChange=NULL;
 
    if (mulSynFast!=NULL) delete[] mulSynFast;
    if (mulSynSlow!=NULL) delete[] mulSynSlow;
    if (managerRuntimeData.connIdsPreIdx!=NULL) delete[] managerRuntimeData.connIdsPreIdx;
    mulSynFast=NULL; mulSynSlow=NULL; managerRuntimeData.connIdsPreIdx=NULL;
 
    if (managerRuntimeData.grpIds!=NULL) delete[] managerRuntimeData.grpIds;
    managerRuntimeData.grpIds=NULL;
 
    if (managerRuntimeData.timeTableD2 != NULL) delete [] managerRuntimeData.timeTableD2;
    if (managerRuntimeData.timeTableD1 != NULL) delete [] managerRuntimeData.timeTableD1;
    managerRuntimeData.timeTableD2 = NULL; managerRuntimeData.timeTableD1 = NULL;
 
    if (managerRuntimeData.firingTableD2!=NULL) delete[] managerRuntimeData.firingTableD2;
    if (managerRuntimeData.firingTableD1!=NULL) delete[] managerRuntimeData.firingTableD1;
    //if (managerRuntimeData.firingTableD2!=NULL) CUDA_CHECK_ERRORS(cudaFreeHost(managerRuntimeData.firingTableD2));
    //if (managerRuntimeData.firingTableD1!=NULL) CUDA_CHECK_ERRORS(cudaFreeHost(managerRuntimeData.firingTableD1));
    managerRuntimeData.firingTableD2 = NULL; managerRuntimeData.firingTableD1 = NULL;
 
    if (managerRuntimeData.extFiringTableD2!=NULL) delete[] managerRuntimeData.extFiringTableD2;
    if (managerRuntimeData.extFiringTableD1!=NULL) delete[] managerRuntimeData.extFiringTableD1;
    //if (managerRuntimeData.extFiringTableD2!=NULL) CUDA_CHECK_ERRORS(cudaFreeHost(managerRuntimeData.extFiringTableD2));
    //if (managerRuntimeData.extFiringTableD1!=NULL) CUDA_CHECK_ERRORS(cudaFreeHost(managerRuntimeData.extFiringTableD1));
    managerRuntimeData.extFiringTableD2 = NULL; managerRuntimeData.extFiringTableD1 = NULL;
 
    if (managerRuntimeData.extFiringTableEndIdxD1 != NULL) delete[] managerRuntimeData.extFiringTableEndIdxD1;
    if (managerRuntimeData.extFiringTableEndIdxD2 != NULL) delete[] managerRuntimeData.extFiringTableEndIdxD2;
    //if (managerRuntimeData.extFiringTableEndIdxD1 != NULL) CUDA_CHECK_ERRORS(cudaFreeHost(managerRuntimeData.extFiringTableEndIdxD1));
    //if (managerRuntimeData.extFiringTableEndIdxD2 != NULL) CUDA_CHECK_ERRORS(cudaFreeHost(managerRuntimeData.extFiringTableEndIdxD2));
    managerRuntimeData.extFiringTableEndIdxD1 = NULL; managerRuntimeData.extFiringTableEndIdxD2 = NULL;
}
 
void SNN::resetPoissonNeuron(int netId, int lGrpId, int lNId) {
    assert(lNId < networkConfigs[netId].numN);
    managerRuntimeData.lastSpikeTime[lNId] = MAX_SIMULATION_TIME;
    if (groupConfigs[netId][lGrpId].WithHomeostasis)
        managerRuntimeData.avgFiring[lNId] = 0.0f;
 
    if (groupConfigs[netId][lGrpId].WithSTP) {
        for (int j = 0; j < networkConfigs[netId].maxDelay + 1; j++) { // is of size maxDelay_+1
            int index = STP_BUF_POS(lNId, j, networkConfigs[netId].maxDelay);
            managerRuntimeData.stpu[index] = 0.0f;
            managerRuntimeData.stpx[index] = 1.0f;
        }
    }
}
 
void SNN::resetPropogationBuffer() {
    // FIXME: why 1023?
    spikeBuf->reset(0, 1023);
}
 
//Reset wt, wtChange, pre-firing time values to default values, rewritten to
//integrate changes between JMN and MDR -- KDC
//if changeWeights is false, we should keep the values of the weights as they currently
//are but we should be able to change them to plastic or fixed synapses. -- KDC
// FIXME: imlement option of resetting weights
void SNN::resetSynapse(int netId, bool changeWeights) {
    memset(managerRuntimeData.wtChange, 0, sizeof(float) * networkConfigs[netId].numPreSynNet); // reset the synaptic derivatives
 
    for (int syn = 0; syn < networkConfigs[netId].numPreSynNet; syn++)
        managerRuntimeData.synSpikeTime[syn] = MAX_SIMULATION_TIME; // reset the spike time of each syanpse
}
 
void SNN::resetTimeTable() {
    memset(managerRuntimeData.timeTableD2, 0, sizeof(int) * (1000 + glbNetworkConfig.maxDelay + 1));
    memset(managerRuntimeData.timeTableD1, 0, sizeof(int) * (1000 + glbNetworkConfig.maxDelay + 1));
}
 
void SNN::resetFiringTable() {
    memset(managerRuntimeData.firingTableD2, 0, sizeof(int) * managerRTDSize.maxMaxSpikeD2);
    memset(managerRuntimeData.firingTableD1, 0, sizeof(int) * managerRTDSize.maxMaxSpikeD1);
    memset(managerRuntimeData.extFiringTableEndIdxD2, 0, sizeof(int) * managerRTDSize.maxNumGroups);
    memset(managerRuntimeData.extFiringTableEndIdxD1, 0, sizeof(int) * managerRTDSize.maxNumGroups);
    memset(managerRuntimeData.extFiringTableD2, 0, sizeof(int*) * managerRTDSize.maxNumGroups);
    memset(managerRuntimeData.extFiringTableD1, 0, sizeof(int*) * managerRTDSize.maxNumGroups);
}
 
void SNN::resetSpikeCnt(int gGrpId) {
    assert(gGrpId >= ALL);
 
    if (gGrpId == ALL) {
        #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
            pthread_t threads[numCores + 1]; // 1 additional array size if numCores == 0, it may work though bad practice
            cpu_set_t cpus;
            ThreadStruct argsThreadRoutine[numCores + 1]; // same as above, +1 array size
            int threadCount = 0;
        #endif
 
        for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
            if (!groupPartitionLists[netId].empty()) {
                if (netId < CPU_RUNTIME_BASE) // GPU runtime
                    resetSpikeCnt_GPU(netId, ALL);
                else{ // CPU runtime
                    #if defined(WIN32) || defined(WIN64) || defined(__APPLE__)
                        resetSpikeCnt_CPU(netId, ALL);
                    #else // Linux or MAC
                        pthread_attr_t attr;
                        pthread_attr_init(&attr);
                        CPU_ZERO(&cpus);
                        CPU_SET(threadCount%NUM_CPU_CORES, &cpus);
                        pthread_attr_setaffinity_np(&attr, sizeof(cpu_set_t), &cpus);
 
                        argsThreadRoutine[threadCount].snn_pointer = this;
                        argsThreadRoutine[threadCount].netId = netId;
                        argsThreadRoutine[threadCount].lGrpId = ALL;
                        argsThreadRoutine[threadCount].startIdx = 0;
                        argsThreadRoutine[threadCount].endIdx = 0;
                        argsThreadRoutine[threadCount].GtoLOffset = 0;
 
                        pthread_create(&threads[threadCount], &attr, &SNN::helperResetSpikeCnt_CPU, (void*)&argsThreadRoutine[threadCount]);
                        pthread_attr_destroy(&attr);
                        threadCount++;
                    #endif
                }
            }
        }
 
        #if !defined(WIN32) && !defined(WIN64) && !defined(__APPLE__) // Linux or MAC
            // join all the threads
            for (int i=0; i<threadCount; i++){
                pthread_join(threads[i], NULL);
            }
        #endif
    }
    else {
        int netId = groupConfigMDMap[gGrpId].netId;
        int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
 
        if (netId < CPU_RUNTIME_BASE) // GPU runtime
            resetSpikeCnt_GPU(netId, lGrpId);
        else // CPU runtime
            resetSpikeCnt_CPU(netId, lGrpId);
    }
}
 
 
inline SynInfo SNN::SET_CONN_ID(int nId, int sId, int grpId) {
    if (grpId > GROUP_ID_MASK) {
        KERNEL_ERROR("Error: Group Id (%d) exceeds maximum limit (%d)", grpId, GROUP_ID_MASK);
        exitSimulation(ID_OVERFLOW_ERROR);
    }
 
    SynInfo synInfo;
    //p.postId = (((sid)<<CONN_SYN_NEURON_BITS)+((nid)&CONN_SYN_NEURON_MASK));
    //p.grpId  = grpId;
    synInfo.gsId = ((grpId << NUM_SYNAPSE_BITS) | sId);
    synInfo.nId = nId;
 
    return synInfo;
}
 
 
void SNN::setGrpTimeSlice(int gGrpId, int timeSlice) {
    if (gGrpId == ALL) {
        for(int grpId = 0; grpId < numGroups; grpId++) {
            if (groupConfigMap[grpId].isSpikeGenerator)
                setGrpTimeSlice(grpId, timeSlice);
        }
    } else {
        assert((timeSlice > 0 ) && (timeSlice <= MAX_TIME_SLICE));
        // the group should be poisson spike generator group
        groupConfigMDMap[gGrpId].currTimeSlice = timeSlice;
    }
}
 
// method to set const member randSeed_
int SNN::setRandSeed(int seed) {
    if (seed<0)
        return time(NULL);
    else if(seed==0)
        return 123;
    else
        return seed;
}
 
void SNN::fillSpikeGenBits(int netId) {
    SpikeBuffer::SpikeIterator spikeBufIter;
    SpikeBuffer::SpikeIterator spikeBufIterEnd = spikeBuf->back();
 
    // Covert spikes stored in spikeBuffer to SpikeGenBit
    for (spikeBufIter = spikeBuf->front(); spikeBufIter != spikeBufIterEnd; ++spikeBufIter) {
        // get the global neuron id and group id for this particular spike
        int gGrpId = spikeBufIter->grpId;
 
        if (groupConfigMDMap[gGrpId].netId == netId) {
            int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
            int lNId = spikeBufIter->neurId /* gNId */ + groupConfigMDMap[gGrpId].GtoLOffset;
 
            // add spike to spikeGentBit
            assert(groupConfigMap[gGrpId].isSpikeGenerator == true);
 
            int nIdPos = (lNId - groupConfigs[netId][lGrpId].lStartN + groupConfigs[netId][lGrpId].Noffset);
            int nIdBitPos = nIdPos % 32;
            int nIdIndex = nIdPos / 32;
 
            assert(nIdIndex < (networkConfigs[netId].numNSpikeGen / 32 + 1));
 
            managerRuntimeData.spikeGenBits[nIdIndex] |= (1 << nIdBitPos);
        }
    }
}
 
void SNN::startTiming() { prevExecutionTime = cumExecutionTime; }
void SNN::stopTiming() {
    executionTime += (cumExecutionTime - prevExecutionTime);
    prevExecutionTime = cumExecutionTime;
}
 
// enters testing phase
// in testing, no weight changes can be made, allowing you to evaluate learned weights, etc.
void SNN::startTesting(bool shallUpdateWeights) {
    // because this can be called at any point in time, if we're off the 1-second grid, we want to make
    // sure to apply the accumulated weight changes to the weight matrix
    // but we don't reset the wt update interval counter
    if (shallUpdateWeights && !sim_in_testing) {
        // careful: need to temporarily adjust stdpScaleFactor to make this right
        if (wtANDwtChangeUpdateIntervalCnt_) {
            float storeScaleSTDP = stdpScaleFactor_;
            stdpScaleFactor_ = 1.0f/wtANDwtChangeUpdateIntervalCnt_;
 
            updateWeights();
 
            stdpScaleFactor_ = storeScaleSTDP;
        }
    }
 
    sim_in_testing = true;
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            networkConfigs[netId].sim_in_testing = true;
            updateNetworkConfig(netId); // update networkConfigRT struct (|TODO copy only a single boolean)
        }
    }
}
 
// exits testing phase
void SNN::stopTesting() {
    sim_in_testing = false;
 
    for (int netId = 0; netId < MAX_NET_PER_SNN; netId++) {
        if (!groupPartitionLists[netId].empty()) {
            networkConfigs[netId].sim_in_testing = false;
            updateNetworkConfig(netId); // update networkConfigRT struct (|TODO copy only a single boolean)
        }
    }
}
 
void SNN::updateConnectionMonitor(short int connId) {
    for (int monId=0; monId<numConnectionMonitor; monId++) {
        if (connId==ALL || connMonCoreList[monId]->getConnectId()==connId) {
            int timeInterval = connMonCoreList[monId]->getUpdateTimeIntervalSec();
            if (timeInterval==1 || timeInterval>1 && (getSimTime()%timeInterval)==0) {
                // this ConnectionMonitor wants periodic recording
                connMonCoreList[monId]->writeConnectFileSnapshot(simTime,
                    getWeightMatrix2D(connMonCoreList[monId]->getConnectId()));
            }
        }
    }
}
 
// FIXME: modify this for multi-GPUs
std::vector< std::vector<float> > SNN::getWeightMatrix2D(short int connId) {
    assert(connId > ALL); // ALL == -1
    std::vector< std::vector<float> > wtConnId;
 
    int grpIdPre = connectConfigMap[connId].grpSrc;
    int grpIdPost = connectConfigMap[connId].grpDest;
 
    int netIdPost = groupConfigMDMap[grpIdPost].netId;
    int lGrpIdPost = groupConfigMDMap[grpIdPost].lGrpId;
 
    // init weight matrix with right dimensions
    for (int i = 0; i < groupConfigMap[grpIdPre].numN; i++) {
        std::vector<float> wtSlice;
        for (int j = 0; j < groupConfigMap[grpIdPost].numN; j++) {
            wtSlice.push_back(NAN);
        }
        wtConnId.push_back(wtSlice);
    }
 
    // copy the weights for a given post-group from device
    // \TODO: check if the weights for this grpIdPost have already been copied
    // \TODO: even better, but tricky because of ordering, make copyWeightState connection-based
 
    assert(grpIdPost > ALL); // ALL == -1
 
    // Note, copyWeightState() also copies pre-connections information (e.g., Npre, Npre_plastic, cumulativePre, and preSynapticIds)
    fetchWeightState(netIdPost, lGrpIdPost);
    fetchConnIdsLookupArray(netIdPost);
 
    for (int lNIdPost = groupConfigs[netIdPost][lGrpIdPost].lStartN; lNIdPost <= groupConfigs[netIdPost][lGrpIdPost].lEndN; lNIdPost++) {
        unsigned int pos_ij = managerRuntimeData.cumulativePre[lNIdPost];
        for (int i = 0; i < managerRuntimeData.Npre[lNIdPost]; i++, pos_ij++) {
            // skip synapses that belong to a different connection ID
            if (managerRuntimeData.connIdsPreIdx[pos_ij] != connId) //connInfo->connId)
                continue;
 
            // find pre-neuron ID and update ConnectionMonitor container
            int lNIdPre = GET_CONN_NEURON_ID(managerRuntimeData.preSynapticIds[pos_ij]);
            int lGrpIdPre = GET_CONN_GRP_ID(managerRuntimeData.preSynapticIds[pos_ij]);
            wtConnId[lNIdPre - groupConfigs[netIdPost][lGrpIdPre].lStartN][lNIdPost - groupConfigs[netIdPost][lGrpIdPost].lStartN] =
                fabs(managerRuntimeData.wt[pos_ij]);
        }
    }
 
    return wtConnId;
}
 
void SNN::updateGroupMonitor(int gGrpId) {
    // don't continue if no group monitors in the network
    if (!numGroupMonitor)
        return;
 
    if (gGrpId == ALL) {
        for (int gGrpId = 0; gGrpId < numGroups; gGrpId++)
            updateGroupMonitor(gGrpId);
    } else {
        int netId = groupConfigMDMap[gGrpId].netId;
        int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
        // update group monitor of a specific group
        // find index in group monitor arrays
        int monitorId = groupConfigMDMap[gGrpId].groupMonitorId;
 
        // don't continue if no group monitor enabled for this group
        if (monitorId < 0) return;
 
        // find last update time for this group
        GroupMonitorCore* grpMonObj = groupMonCoreList[monitorId];
        int lastUpdate = grpMonObj->getLastUpdated();
 
        // don't continue if time interval is zero (nothing to update)
        if (getSimTime() - lastUpdate <= 0)
            return;
 
        if (getSimTime() - lastUpdate > 1000)
            KERNEL_ERROR("updateGroupMonitor(grpId=%d) must be called at least once every second", gGrpId);
 
        // copy the group status (neuromodulators) to the manager runtime
        fetchGroupState(netId, lGrpId);
 
        // find the time interval in which to update group status
        // usually, we call updateGroupMonitor once every second, so the time interval is [0,1000)
        // however, updateGroupMonitor can be called at any time t \in [0,1000)... so we can have the cases
        // [0,t), [t,1000), and even [t1, t2)
        int numMsMin = lastUpdate % 1000; // lower bound is given by last time we called update
        int numMsMax = getSimTimeMs(); // upper bound is given by current time
        if (numMsMax == 0)
            numMsMax = 1000; // special case: full second
        assert(numMsMin < numMsMax);
 
        // current time is last completed second in milliseconds (plus t to be added below)
        // special case is after each completed second where !getSimTimeMs(): here we look 1s back
        int currentTimeSec = getSimTimeSec();
        if (!getSimTimeMs())
            currentTimeSec--;
 
        // save current time as last update time
        grpMonObj->setLastUpdated(getSimTime());
 
        // prepare fast access
        FILE* grpFileId = groupMonCoreList[monitorId]->getGroupFileId();
        bool writeGroupToFile = grpFileId != NULL;
        bool writeGroupToArray = grpMonObj->isRecording();
        float data;
 
        // Read one peice of data at a time from the buffer and put the data to an appopriate monitor buffer. Later the user
        // may need need to dump these group status data to an output file
        for(int t = numMsMin; t < numMsMax; t++) {
            // fetch group status data, support dopamine concentration currently
            data = managerRuntimeData.grpDABuffer[lGrpId * 1000 + t];
 
            // current time is last completed second plus whatever is leftover in t
            int time = currentTimeSec * 1000 + t;
 
            if (writeGroupToFile) {
                // TODO: write to group status file
            }
 
            if (writeGroupToArray) {
                grpMonObj->pushData(time, data);
            }
        }
 
        if (grpFileId!=NULL) // flush group status file
            fflush(grpFileId);
    }
}
 
// FIXME: wrong to use groupConfigs[0]
void SNN::userDefinedSpikeGenerator(int gGrpId) {
    // \FIXME this function is a mess
    SpikeGeneratorCore* spikeGenFunc = groupConfigMap[gGrpId].spikeGenFunc;
    int netId = groupConfigMDMap[gGrpId].netId;
    int timeSlice = groupConfigMDMap[gGrpId].currTimeSlice;
    int currTime = simTime;
    bool done;
 
    fetchLastSpikeTime(netId);
 
    for(int gNId = groupConfigMDMap[gGrpId].gStartN; gNId <= groupConfigMDMap[gGrpId].gEndN; gNId++) {
        // start the time from the last time it spiked, that way we can ensure that the refractory period is maintained
        int lNId = gNId + groupConfigMDMap[gGrpId].GtoLOffset;
        int nextTime = managerRuntimeData.lastSpikeTime[lNId];
        if (nextTime == MAX_SIMULATION_TIME)
            nextTime = 0;
 
        // the end of the valid time window is either the length of the scheduling time slice from now (because that
        // is the max of the allowed propagated buffer size) or simply the end of the simulation
        int endOfTimeWindow = std::min(currTime+timeSlice, simTimeRunStop);
 
        done = false;
        while (!done) {
            // generate the next spike time (nextSchedTime) from the nextSpikeTime callback
            int nextSchedTime = spikeGenFunc->nextSpikeTime(this, gGrpId, gNId - groupConfigMDMap[gGrpId].gStartN, currTime, nextTime, endOfTimeWindow);
 
            // the generated spike time is valid only if:
            // - it has not been scheduled before (nextSchedTime > nextTime)
            //    - but careful: we would drop spikes at t=0, because we cannot initialize nextTime to -1...
            // - it is within the scheduling time slice (nextSchedTime < endOfTimeWindow)
            // - it is not in the past (nextSchedTime >= currTime)
            if ((nextSchedTime==0 || nextSchedTime>nextTime) && nextSchedTime<endOfTimeWindow && nextSchedTime>=currTime) {
//              fprintf(stderr,"%u: spike scheduled for %d at %u\n",currTime, i-groupConfigs[0][grpId].StartN,nextSchedTime);
                // scheduled spike...
                // \TODO CPU mode does not check whether the same AER event has been scheduled before (bug #212)
                // check how GPU mode does it, then do the same here.
                nextTime = nextSchedTime;
                spikeBuf->schedule(gNId, gGrpId, nextTime - currTime);
            } else {
                done = true;
            }
        }
    }
}
 
void SNN::generateUserDefinedSpikes() {
    for(int gGrpId = 0; gGrpId < numGroups; gGrpId++) {
        if (groupConfigMap[gGrpId].isSpikeGenerator) {
            // This evaluation is done to check if its time to get new set of spikes..
            // check whether simTime has advance more than the current time slice, in which case we need to schedule
            // spikes for the next time slice
            // we always have to run this the first millisecond of a new runNetwork call; that is,
            // when simTime==simTimeRunStart
            if(((simTime - groupConfigMDMap[gGrpId].sliceUpdateTime) >= groupConfigMDMap[gGrpId].currTimeSlice || simTime == simTimeRunStart)) {
                int timeSlice = groupConfigMDMap[gGrpId].currTimeSlice;
                groupConfigMDMap[gGrpId].sliceUpdateTime = simTime;
 
                // we dont generate any poisson spike if during the
                // current call we might exceed the maximum 32 bit integer value
                if ((simTime + timeSlice) == MAX_SIMULATION_TIME || (simTime + timeSlice) < 0)
                    return;
 
                if (groupConfigMap[gGrpId].spikeGenFunc != NULL) {
                    userDefinedSpikeGenerator(gGrpId);
                }
            }
        }
    }
}
 
void SNN::allocateManagerSpikeTables() {
    managerRuntimeData.firingTableD2 = new int[managerRTDSize.maxMaxSpikeD2];
    managerRuntimeData.firingTableD1 = new int[managerRTDSize.maxMaxSpikeD1];
    managerRuntimeData.extFiringTableEndIdxD2 = new int[managerRTDSize.maxNumGroups];
    managerRuntimeData.extFiringTableEndIdxD1 = new int[managerRTDSize.maxNumGroups];
    managerRuntimeData.extFiringTableD2 = new int*[managerRTDSize.maxNumGroups];
    managerRuntimeData.extFiringTableD1 = new int*[managerRTDSize.maxNumGroups];
 
    //CUDA_CHECK_ERRORS(cudaMallocHost(&managerRuntimeData.firingTableD2, sizeof(int) * managerRTDSize.maxMaxSpikeD2));
    //CUDA_CHECK_ERRORS(cudaMallocHost(&managerRuntimeData.firingTableD1, sizeof(int) * managerRTDSize.maxMaxSpikeD1));
    //CUDA_CHECK_ERRORS(cudaMallocHost(&managerRuntimeData.extFiringTableEndIdxD2, sizeof(int) * managerRTDSize.maxNumGroups));
    //CUDA_CHECK_ERRORS(cudaMallocHost(&managerRuntimeData.extFiringTableEndIdxD1, sizeof(int) * managerRTDSize.maxNumGroups));
    //CUDA_CHECK_ERRORS(cudaMallocHost(&managerRuntimeData.extFiringTableD2, sizeof(int*) * managerRTDSize.maxNumGroups));
    //CUDA_CHECK_ERRORS(cudaMallocHost(&managerRuntimeData.extFiringTableD1, sizeof(int*) * managerRTDSize.maxNumGroups));
    resetFiringTable();
 
    managerRuntimeData.timeTableD2 = new unsigned int[TIMING_COUNT];
    managerRuntimeData.timeTableD1 = new unsigned int[TIMING_COUNT];
    resetTimeTable();
}
 
// updates simTime, returns true when new second started
bool SNN::updateTime() {
    bool finishedOneSec = false;
 
    // done one second worth of simulation
    // update relevant parameters...now
    if(++simTimeMs == 1000) {
        simTimeMs = 0;
        simTimeSec++;
        finishedOneSec = true;
    }
 
    simTime++;
    if(simTime == MAX_SIMULATION_TIME || simTime < 0){
        // reached the maximum limit of the simulation time using 32 bit value...
        KERNEL_WARN("Maximum Simulation Time Reached...Resetting simulation time");
    }
 
    return finishedOneSec;
}
 
// FIXME: modify this for multi-GPUs
void SNN::updateSpikeMonitor(int gGrpId) {
    // don't continue if no spike monitors in the network
    if (!numSpikeMonitor)
        return;
 
    if (gGrpId == ALL) {
        for (int gGrpId = 0; gGrpId < numGroups; gGrpId++)
            updateSpikeMonitor(gGrpId);
    } else {
        int netId = groupConfigMDMap[gGrpId].netId;
        int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
        // update spike monitor of a specific group
        // find index in spike monitor arrays
        int monitorId = groupConfigMDMap[gGrpId].spikeMonitorId;
 
        // don't continue if no spike monitor enabled for this group
        if (monitorId < 0) return;
 
        // find last update time for this group
        SpikeMonitorCore* spkMonObj = spikeMonCoreList[monitorId];
        long int lastUpdate = spkMonObj->getLastUpdated();
 
        // don't continue if time interval is zero (nothing to update)
        if ( ((long int)getSimTime()) - lastUpdate <= 0)
            return;
 
        if ( ((long int)getSimTime()) - lastUpdate > 1000)
            KERNEL_ERROR("updateSpikeMonitor(grpId=%d) must be called at least once every second",gGrpId);
 
        // AER buffer max size warning here.
        // Because of C++ short-circuit evaluation, the last condition should not be evaluated
        // if the previous conditions are false.
        if (spkMonObj->getAccumTime() > LONG_SPIKE_MON_DURATION \
                && this->getGroupNumNeurons(gGrpId) > LARGE_SPIKE_MON_GRP_SIZE \
                && spkMonObj->isBufferBig()){
            // change this warning message to correct message
            KERNEL_WARN("updateSpikeMonitor(grpId=%d) is becoming very large. (>%lu MB)",gGrpId,(long int) MAX_SPIKE_MON_BUFFER_SIZE/1024 );// make this better
            KERNEL_WARN("Reduce the cumulative recording time (currently %lu minutes) or the group size (currently %d) to avoid this.",spkMonObj->getAccumTime()/(1000*60),this->getGroupNumNeurons(gGrpId));
        }
 
        // copy the neuron firing information to the manager runtime
        fetchSpikeTables(netId);
        fetchGrpIdsLookupArray(netId);
 
        // find the time interval in which to update spikes
        // usually, we call updateSpikeMonitor once every second, so the time interval is [0,1000)
        // however, updateSpikeMonitor can be called at any time t \in [0,1000)... so we can have the cases
        // [0,t), [t,1000), and even [t1, t2)
        int numMsMin = lastUpdate % 1000; // lower bound is given by last time we called update
        int numMsMax = getSimTimeMs(); // upper bound is given by current time
        if (numMsMax == 0)
            numMsMax = 1000; // special case: full second
        assert(numMsMin < numMsMax);
 
        // current time is last completed second in milliseconds (plus t to be added below)
        // special case is after each completed second where !getSimTimeMs(): here we look 1s back
        int currentTimeSec = getSimTimeSec();
        if (!getSimTimeMs())
            currentTimeSec--;
 
        // save current time as last update time
        spkMonObj->setLastUpdated( (long int)getSimTime() );
 
        // prepare fast access
        FILE* spkFileId = spikeMonCoreList[monitorId]->getSpikeFileId();
        bool writeSpikesToFile = spkFileId != NULL;
        bool writeSpikesToArray = spkMonObj->getMode()==AER && spkMonObj->isRecording();
 
        // Read one spike at a time from the buffer and put the spikes to an appopriate monitor buffer. Later the user
        // may need need to dump these spikes to an output file
        for (int k = 0; k < 2; k++) {
            unsigned int* timeTablePtr = (k == 0) ? managerRuntimeData.timeTableD2 : managerRuntimeData.timeTableD1;
            int* fireTablePtr = (k == 0) ? managerRuntimeData.firingTableD2 : managerRuntimeData.firingTableD1;
            for(int t = numMsMin; t < numMsMax; t++) {
                for(int i = timeTablePtr[t + glbNetworkConfig.maxDelay]; i < timeTablePtr[t + glbNetworkConfig.maxDelay + 1]; i++) {
                    // retrieve the neuron id
                    int lNId = fireTablePtr[i];
 
                    // make sure neuron belongs to currently relevant group
                    int this_grpId = managerRuntimeData.grpIds[lNId];
                    if (this_grpId != lGrpId)
                        continue;
 
                    // adjust nid to be 0-indexed for each group
                    // this way, if a group has 10 neurons, their IDs in the spike file and spike monitor will be
                    // indexed from 0..9, no matter what their real nid is
                    int nId = lNId - groupConfigs[netId][lGrpId].lStartN;
                    assert(nId >= 0);
 
                    // current time is last completed second plus whatever is leftover in t
                    int time = currentTimeSec * 1000 + t;
 
                    if (writeSpikesToFile) {
                        int cnt;
                        cnt = fwrite(&time, sizeof(int), 1, spkFileId); assert(cnt==1);
                        cnt = fwrite(&nId,  sizeof(int), 1, spkFileId); assert(cnt==1);
                    }
 
                    if (writeSpikesToArray) {
                        spkMonObj->pushAER(time, nId);
                    }
                }
            }
        }
 
        if (spkFileId!=NULL) // flush spike file
            fflush(spkFileId);
    }
}
 
// FIXME: modify this for multi-GPUs
void SNN::updateNeuronMonitor(int gGrpId) {
    // don't continue if no neuron monitors in the network
    if (!numNeuronMonitor)
        return;
 
    //printf("The global group id is: %i\n", gGrpId);
 
    if (gGrpId == ALL) {
        for (int gGrpId = 0; gGrpId < numGroups; gGrpId++)
            updateNeuronMonitor(gGrpId);
    }
    else {
        //printf("UpdateNeuronMonitor is being executed!\n");
        int netId = groupConfigMDMap[gGrpId].netId;
        int lGrpId = groupConfigMDMap[gGrpId].lGrpId;
        // update neuron monitor of a specific group
        // find index in neuron monitor arrays
        int monitorId = groupConfigMDMap[gGrpId].neuronMonitorId;
 
        // don't continue if no spike monitor enabled for this group
        if (monitorId < 0) return;
 
        // find last update time for this group
        NeuronMonitorCore* nrnMonObj = neuronMonCoreList[monitorId];
        long int lastUpdate = nrnMonObj->getLastUpdated();
 
        // don't continue if time interval is zero (nothing to update)
        if (((long int)getSimTime()) - lastUpdate <= 0)
            return;
 
        if (((long int)getSimTime()) - lastUpdate > 1000)
            KERNEL_ERROR("updateNeuronMonitor(grpId=%d) must be called at least once every second", gGrpId);
 
        // AER buffer max size warning here.
        // Because of C++ short-circuit evaluation, the last condition should not be evaluated
        // if the previous conditions are false.
 
        /*if (nrnMonObj->getAccumTime() > LONG_NEURON_MON_DURATION \
            && this->getGroupNumNeurons(gGrpId) > LARGE_NEURON_MON_GRP_SIZE \
            && nrnMonObj->isBufferBig()) {
            // change this warning message to correct message
            KERNEL_WARN("updateNeuronMonitor(grpId=%d) is becoming very large. (>%lu MB)", gGrpId, (long int)MAX_NEURON_MON_BUFFER_SIZE / 1024);// make this better
            KERNEL_WARN("Reduce the cumulative recording time (currently %lu minutes) or the group size (currently %d) to avoid this.", nrnMonObj->getAccumTime() / (1000 * 60), this->getGroupNumNeurons(gGrpId));
        }*/
 
        // copy the neuron information to manager runtime
        fetchNeuronStateBuffer(netId, lGrpId);
 
        // find the time interval in which to update neuron state info
        // usually, we call updateNeuronMonitor once every second, so the time interval is [0,1000)
        // however, updateNeuronMonitor can be called at any time t \in [0,1000)... so we can have the cases
        // [0,t), [t,1000), and even [t1, t2)
        int numMsMin = lastUpdate % 1000; // lower bound is given by last time we called update
        int numMsMax = getSimTimeMs(); // upper bound is given by current time
        if (numMsMax == 0)
            numMsMax = 1000; // special case: full second
        assert(numMsMin < numMsMax);
        //KERNEL_INFO("lastUpdate: %d -- numMsMin: %d -- numMsMax: %d", lastUpdate, numMsMin, numMsMax);
 
        // current time is last completed second in milliseconds (plus t to be added below)
        // special case is after each completed second where !getSimTimeMs(): here we look 1s back
        int currentTimeSec = getSimTimeSec();
        if (!getSimTimeMs())
            currentTimeSec--;
 
        // save current time as last update time
        nrnMonObj->setLastUpdated((long int)getSimTime());
 
        // prepare fast access
        FILE* nrnFileId = neuronMonCoreList[monitorId]->getNeuronFileId();
        bool writeNeuronStateToFile = nrnFileId != NULL;
        bool writeNeuronStateToArray = nrnMonObj->isRecording();
 
        // Read one neuron state value at a time from the buffer and put the neuron state values to an appopriate monitor buffer.
        // Later the user may need need to dump these neuron state values to an output file
        //printf("The numMsMin is: %i; and numMsMax is: %i\n", numMsMin, numMsMax);
        for (int t = numMsMin; t < numMsMax; t++) {
            //printf("The lStartN is: %i; and lEndN is: %i\n", groupConfigs[netId][lGrpId].lStartN, groupConfigs[netId][lGrpId].lEndN);
            for (int lNId = groupConfigs[netId][lGrpId].lStartN; lNId <= groupConfigs[netId][lGrpId].lEndN; lNId++) {
                float v, u, I;
 
                // make sure neuron belongs to currently relevant group
                int this_grpId = managerRuntimeData.grpIds[lNId];
                if (this_grpId != lGrpId)
                    continue;
 
                // adjust nid to be 0-indexed for each group
                // this way, if a group has 10 neurons, their IDs in the spike file and spike monitor will be
                // indexed from 0..9, no matter what their real nid is
                int nId = lNId - groupConfigs[netId][lGrpId].lStartN;
                assert(nId >= 0);
 
                int idxBase = networkConfigs[netId].numGroups * MAX_NEURON_MON_GRP_SZIE * t + lGrpId * MAX_NEURON_MON_GRP_SZIE;
                v = managerRuntimeData.nVBuffer[idxBase + nId];
                u = managerRuntimeData.nUBuffer[idxBase + nId];
                I = managerRuntimeData.nIBuffer[idxBase + nId];
 
                //printf("Voltage recorded is: %f\n", v);
 
                // current time is last completed second plus whatever is leftover in t
                int time = currentTimeSec * 1000 + t;
 
                //KERNEL_INFO("t: %d -- time: %d --base: %d -- nId: %d -- v: %f -- u: %f, --I: %f", t, time, idxBase + nId, nId, v, u, I);
 
                // WRITE TO A TEXT FILE INSTEAD OF BINARY
                if (writeNeuronStateToFile) {
                    //KERNEL_INFO("Save to file");
                    int cnt;
                    cnt = fwrite(&time, sizeof(int), 1, nrnFileId); assert(cnt == 1);
                    cnt = fwrite(&nId, sizeof(int), 1, nrnFileId); assert(cnt == 1);
                    cnt = fwrite(&v, sizeof(float), 1, nrnFileId); assert(cnt == 1);
                    cnt = fwrite(&u, sizeof(float), 1, nrnFileId); assert(cnt == 1);
                    cnt = fwrite(&I, sizeof(float), 1, nrnFileId); assert(cnt == 1);
                }
 
                if (writeNeuronStateToArray) {
                    //KERNEL_INFO("Save to array");
                    nrnMonObj->pushNeuronState(nId, v, u, I);
                }
            }
        }
 
        if (nrnFileId != NULL) // flush neuron state file
            fflush(nrnFileId);
    }
}
 
// FIXME: update summary format for multiGPUs
void SNN::printSimSummary() {
    float etime;
 
    // FIXME: measure total execution time, and GPU excution time
    stopTiming();
    etime = executionTime;
 
    fetchNetworkSpikeCount();
 
    KERNEL_INFO("\n");
    KERNEL_INFO("********************    Simulation Summary      ***************************");
 
    KERNEL_INFO("Network Parameters: \tnumNeurons = %d (numNExcReg:numNInhReg = %2.1f:%2.1f)",
        glbNetworkConfig.numN, 100.0 * glbNetworkConfig.numNExcReg / glbNetworkConfig.numN, 100.0 * glbNetworkConfig.numNInhReg / glbNetworkConfig.numN);
    KERNEL_INFO("\t\t\tnumSynapses = %d", glbNetworkConfig.numSynNet);
    KERNEL_INFO("\t\t\tmaxDelay = %d", glbNetworkConfig.maxDelay);
    KERNEL_INFO("Simulation Mode:\t%s",sim_with_conductances?"COBA":"CUBA");
    KERNEL_INFO("Random Seed:\t\t%d", randSeed_);
    KERNEL_INFO("Timing:\t\t\tModel Simulation Time = %lld sec", (unsigned long long)simTimeSec);
    KERNEL_INFO("\t\t\tActual Execution Time = %4.2f sec", etime/1000.0f);
    KERNEL_INFO("Average Firing Rate:\t2+ms delay = %3.3f Hz",
        glbNetworkConfig.numN2msDelay > 0 ? managerRuntimeData.spikeCountD2 / (1.0 * simTimeSec * glbNetworkConfig.numN2msDelay) : 0.0f);
    KERNEL_INFO("\t\t\t1ms delay = %3.3f Hz",
        glbNetworkConfig.numN1msDelay > 0 ? managerRuntimeData.spikeCountD1 / (1.0 * simTimeSec * glbNetworkConfig.numN1msDelay) : 0.0f);
    KERNEL_INFO("\t\t\tOverall = %3.3f Hz", managerRuntimeData.spikeCount / (1.0 * simTimeSec * glbNetworkConfig.numN));
    KERNEL_INFO("Overall Spike Count Transferred:");
    KERNEL_INFO("\t\t\t2+ms delay = %d", managerRuntimeData.spikeCountExtRxD2);
    KERNEL_INFO("\t\t\t1ms delay = %d", managerRuntimeData.spikeCountExtRxD1);
    KERNEL_INFO("Overall Spike Count:\t2+ms delay = %d", managerRuntimeData.spikeCountD2);
    KERNEL_INFO("\t\t\t1ms delay = %d", managerRuntimeData.spikeCountD1);
    KERNEL_INFO("\t\t\tTotal = %d", managerRuntimeData.spikeCount);
    KERNEL_INFO("*********************************************************************************\n");
}
 
//------------------------------ legacy code --------------------------------//
 
// We parallelly cleanup the postSynapticIds array to minimize any other wastage in that array by compacting the store
// Appropriate alignment specified by ALIGN_COMPACTION macro is used to ensure some level of alignment (if necessary)
//void SNN::compactConnections() {
//  unsigned int* tmp_cumulativePost = new unsigned int[numN];
//  unsigned int* tmp_cumulativePre  = new unsigned int[numN];
//  unsigned int lastCnt_pre         = 0;
//  unsigned int lastCnt_post        = 0;
//
//  tmp_cumulativePost[0]   = 0;
//  tmp_cumulativePre[0]    = 0;
//
//  for(int i=1; i < numN; i++) {
//      lastCnt_post = tmp_cumulativePost[i-1]+managerRuntimeData.Npost[i-1]; //position of last pointer
//      lastCnt_pre  = tmp_cumulativePre[i-1]+managerRuntimeData.Npre[i-1]; //position of last pointer
//      #if COMPACTION_ALIGNMENT_POST
//          lastCnt_post= lastCnt_post + COMPACTION_ALIGNMENT_POST-lastCnt_post%COMPACTION_ALIGNMENT_POST;
//          lastCnt_pre = lastCnt_pre  + COMPACTION_ALIGNMENT_PRE- lastCnt_pre%COMPACTION_ALIGNMENT_PRE;
//      #endif
//      tmp_cumulativePost[i] = lastCnt_post;
//      tmp_cumulativePre[i]  = lastCnt_pre;
//      assert(tmp_cumulativePost[i] <= managerRuntimeData.cumulativePost[i]);
//      assert(tmp_cumulativePre[i]  <= managerRuntimeData.cumulativePre[i]);
//  }
//
//  // compress the post_synaptic array according to the new values of the tmp_cumulative counts....
//  unsigned int tmp_numPostSynNet = tmp_cumulativePost[numN-1]+managerRuntimeData.Npost[numN-1];
//  unsigned int tmp_numPreSynNet  = tmp_cumulativePre[numN-1]+managerRuntimeData.Npre[numN-1];
//  assert(tmp_numPostSynNet <= allocatedPost);
//  assert(tmp_numPreSynNet  <= allocatedPre);
//  assert(tmp_numPostSynNet <= numPostSynNet);
//  assert(tmp_numPreSynNet  <= numPreSynNet);
//  KERNEL_DEBUG("******************");
//  KERNEL_DEBUG("CompactConnection: ");
//  KERNEL_DEBUG("******************");
//  KERNEL_DEBUG("old_postCnt = %d, new_postCnt = %d", numPostSynNet, tmp_numPostSynNet);
//  KERNEL_DEBUG("old_preCnt = %d,  new_postCnt = %d", numPreSynNet,  tmp_numPreSynNet);
//
//  // new buffer with required size + 100 bytes of additional space just to provide limited overflow
//  SynInfo* tmp_postSynapticIds   = new SynInfo[tmp_numPostSynNet+100];
//
//  // new buffer with required size + 100 bytes of additional space just to provide limited overflow
//  SynInfo* tmp_preSynapticIds = new SynInfo[tmp_numPreSynNet+100];
//  float* tmp_wt                   = new float[tmp_numPreSynNet+100];
//  float* tmp_maxSynWt             = new float[tmp_numPreSynNet+100];
//  short int *tmp_cumConnIdPre         = new short int[tmp_numPreSynNet+100];
//  float *tmp_mulSynFast           = new float[numConnections];
//  float *tmp_mulSynSlow           = new float[numConnections];
//
//  // compact synaptic information
//  for(int i=0; i<numN; i++) {
//      assert(tmp_cumulativePost[i] <= managerRuntimeData.cumulativePost[i]);
//      assert(tmp_cumulativePre[i]  <= managerRuntimeData.cumulativePre[i]);
//      for( int j=0; j<managerRuntimeData.Npost[i]; j++) {
//          unsigned int tmpPos = tmp_cumulativePost[i]+j;
//          unsigned int oldPos = managerRuntimeData.cumulativePost[i]+j;
//          tmp_postSynapticIds[tmpPos] = managerRuntimeData.postSynapticIds[oldPos];
//          tmp_SynapticDelay[tmpPos]   = tmp_SynapticDelay[oldPos];
//      }
//      for( int j=0; j<managerRuntimeData.Npre[i]; j++) {
//          unsigned int tmpPos =  tmp_cumulativePre[i]+j;
//          unsigned int oldPos =  managerRuntimeData.cumulativePre[i]+j;
//          tmp_preSynapticIds[tmpPos]  = managerRuntimeData.preSynapticIds[oldPos];
//          tmp_maxSynWt[tmpPos]        = managerRuntimeData.maxSynWt[oldPos];
//          tmp_wt[tmpPos]              = managerRuntimeData.wt[oldPos];
//          tmp_cumConnIdPre[tmpPos]    = managerRuntimeData.connIdsPreIdx[oldPos];
//      }
//  }
//
//  // delete old buffer space
//  delete[] managerRuntimeData.postSynapticIds;
//  managerRuntimeData.postSynapticIds = tmp_postSynapticIds;
//  cpuSnnSz.networkInfoSize -= (sizeof(SynInfo)*numPostSynNet);
//  cpuSnnSz.networkInfoSize += (sizeof(SynInfo)*(tmp_numPostSynNet+100));
//
//  delete[] managerRuntimeData.cumulativePost;
//  managerRuntimeData.cumulativePost  = tmp_cumulativePost;
//
//  delete[] managerRuntimeData.cumulativePre;
//  managerRuntimeData.cumulativePre   = tmp_cumulativePre;
//
//  delete[] managerRuntimeData.maxSynWt;
//  managerRuntimeData.maxSynWt = tmp_maxSynWt;
//  cpuSnnSz.synapticInfoSize -= (sizeof(float)*numPreSynNet);
//  cpuSnnSz.synapticInfoSize += (sizeof(float)*(tmp_numPreSynNet+100));
//
//  delete[] managerRuntimeData.wt;
//  managerRuntimeData.wt = tmp_wt;
//  cpuSnnSz.synapticInfoSize -= (sizeof(float)*numPreSynNet);
//  cpuSnnSz.synapticInfoSize += (sizeof(float)*(tmp_numPreSynNet+100));
//
//  delete[] managerRuntimeData.connIdsPreIdx;
//  managerRuntimeData.connIdsPreIdx = tmp_cumConnIdPre;
//  cpuSnnSz.synapticInfoSize -= (sizeof(short int)*numPreSynNet);
//  cpuSnnSz.synapticInfoSize += (sizeof(short int)*(tmp_numPreSynNet+100));
//
//  // compact connection-centric information
//  for (int i=0; i<numConnections; i++) {
//      tmp_mulSynFast[i] = mulSynFast[i];
//      tmp_mulSynSlow[i] = mulSynSlow[i];
//  }
//  delete[] mulSynFast;
//  delete[] mulSynSlow;
//  mulSynFast = tmp_mulSynFast;
//  mulSynSlow = tmp_mulSynSlow;
//  cpuSnnSz.networkInfoSize -= (2*sizeof(uint8_t)*numPreSynNet);
//  cpuSnnSz.networkInfoSize += (2*sizeof(uint8_t)*(tmp_numPreSynNet+100));
//
//
//  delete[] managerRuntimeData.preSynapticIds;
//  managerRuntimeData.preSynapticIds  = tmp_preSynapticIds;
//  cpuSnnSz.synapticInfoSize -= (sizeof(SynInfo)*numPreSynNet);
//  cpuSnnSz.synapticInfoSize += (sizeof(SynInfo)*(tmp_numPreSynNet+100));
//
//  numPreSynNet    = tmp_numPreSynNet;
//  numPostSynNet   = tmp_numPostSynNet;
//}
 
//The post synaptic connections are sorted based on delay here so that we can reduce storage requirement
//and generation of spike at the post-synaptic side.
//We also create the delay_info array has the delay_start and delay_length parameter
//void SNN::reorganizeDelay()
//{
//  for(int grpId=0; grpId < numGroups; grpId++) {
//      for(int nid=groupConfigs[0][grpId].StartN; nid <= groupConfigs[0][grpId].EndN; nid++) {
//          unsigned int jPos=0;                    // this points to the top of the delay queue
//          unsigned int cumN=managerRuntimeData.cumulativePost[nid];   // cumulativePost[] is unsigned int
//          unsigned int cumDelayStart=0;           // Npost[] is unsigned short
//          for(int td = 0; td < maxDelay_; td++) {
//              unsigned int j=jPos;                // start searching from top of the queue until the end
//              unsigned int cnt=0;                 // store the number of nodes with a delay of td;
//              while(j < managerRuntimeData.Npost[nid]) {
//                  // found a node j with delay=td and we put
//                  // the delay value = 1 at array location td=0;
//                  if(td==(tmp_SynapticDelay[cumN+j]-1)) {
//                      assert(jPos<managerRuntimeData.Npost[nid]);
//                      swapConnections(nid, j, jPos);
//
//                      jPos=jPos+1;
//                      cnt=cnt+1;
//                  }
//                  j=j+1;
//              }
//
//              // update the delay_length and start values...
//              managerRuntimeData.postDelayInfo[nid*(maxDelay_+1)+td].delay_length      = cnt;
//              managerRuntimeData.postDelayInfo[nid*(maxDelay_+1)+td].delay_index_start  = cumDelayStart;
//              cumDelayStart += cnt;
//
//              assert(cumDelayStart <= managerRuntimeData.Npost[nid]);
//          }
//
//          // total cumulative delay should be equal to number of post-synaptic connections at the end of the loop
//          assert(cumDelayStart == managerRuntimeData.Npost[nid]);
//          for(unsigned int j=1; j < managerRuntimeData.Npost[nid]; j++) {
//              unsigned int cumN=managerRuntimeData.cumulativePost[nid]; // cumulativePost[] is unsigned int
//              if( tmp_SynapticDelay[cumN+j] < tmp_SynapticDelay[cumN+j-1]) {
//                  KERNEL_ERROR("Post-synaptic delays not sorted correctly... id=%d, delay[%d]=%d, delay[%d]=%d",
//                      nid, j, tmp_SynapticDelay[cumN+j], j-1, tmp_SynapticDelay[cumN+j-1]);
//                  assert( tmp_SynapticDelay[cumN+j] >= tmp_SynapticDelay[cumN+j-1]);
//              }
//          }
//      }
//  }
//}
 
//void SNN::swapConnections(int nid, int oldPos, int newPos) {
//  unsigned int cumN=managerRuntimeData.cumulativePost[nid];
//
//  // Put the node oldPos to the top of the delay queue
//  SynInfo tmp = managerRuntimeData.postSynapticIds[cumN+oldPos];
//  managerRuntimeData.postSynapticIds[cumN+oldPos]= managerRuntimeData.postSynapticIds[cumN+newPos];
//  managerRuntimeData.postSynapticIds[cumN+newPos]= tmp;
//
//  // Ensure that you have shifted the delay accordingly....
//  uint8_t tmp_delay = tmp_SynapticDelay[cumN+oldPos];
//  tmp_SynapticDelay[cumN+oldPos] = tmp_SynapticDelay[cumN+newPos];
//  tmp_SynapticDelay[cumN+newPos] = tmp_delay;
//
//  // update the pre-information for the postsynaptic neuron at the position oldPos.
//  SynInfo  postInfo = managerRuntimeData.postSynapticIds[cumN+oldPos];
//  int  post_nid = GET_CONN_NEURON_ID(postInfo);
//  int  post_sid = GET_CONN_SYN_ID(postInfo);
//
//  SynInfo* preId    = &(managerRuntimeData.preSynapticIds[managerRuntimeData.cumulativePre[post_nid]+post_sid]);
//  int  pre_nid  = GET_CONN_NEURON_ID((*preId));
//  int  pre_sid  = GET_CONN_SYN_ID((*preId));
//  int  pre_gid  = GET_CONN_GRP_ID((*preId));
//  assert (pre_nid == nid);
//  assert (pre_sid == newPos);
//  *preId = SET_CONN_ID( pre_nid, oldPos, pre_gid);
//
//  // update the pre-information for the postsynaptic neuron at the position newPos
//  postInfo = managerRuntimeData.postSynapticIds[cumN+newPos];
//  post_nid = GET_CONN_NEURON_ID(postInfo);
//  post_sid = GET_CONN_SYN_ID(postInfo);
//
//  preId    = &(managerRuntimeData.preSynapticIds[managerRuntimeData.cumulativePre[post_nid]+post_sid]);
//  pre_nid  = GET_CONN_NEURON_ID((*preId));
//  pre_sid  = GET_CONN_SYN_ID((*preId));
//  pre_gid  = GET_CONN_GRP_ID((*preId));
//  assert (pre_nid == nid);
//  assert (pre_sid == oldPos);
//  *preId = SET_CONN_ID( pre_nid, newPos, pre_gid);
//}
 
// set one specific connection from neuron id 'src' to neuron id 'dest'
//inline void SNN::setConnection(int srcGrp,  int destGrp,  unsigned int src, unsigned int dest, float synWt,
//                                  float maxWt, uint8_t dVal, int connProp, short int connId) {
//  assert(dest<=CONN_SYN_NEURON_MASK);         // total number of neurons is less than 1 million within a GPU
//  assert((dVal >=1) && (dVal <= maxDelay_));
//
//  // adjust sign of weight based on pre-group (negative if pre is inhibitory)
//  synWt = isExcitatoryGroup(srcGrp) ? fabs(synWt) : -1.0*fabs(synWt);
//  maxWt = isExcitatoryGroup(srcGrp) ? fabs(maxWt) : -1.0*fabs(maxWt);
//
//  // we have exceeded the number of possible connection for one neuron
//  if(managerRuntimeData.Npost[src] >= groupConfigs[0][srcGrp].numPostSynapses)    {
//      KERNEL_ERROR("setConnection(%d (Grp=%s), %d (Grp=%s), %f, %d)", src, groupInfo[srcGrp].Name.c_str(),
//                  dest, groupInfo[destGrp].Name.c_str(), synWt, dVal);
//      KERNEL_ERROR("Large number of postsynaptic connections established (%d), max for this group %d.", managerRuntimeData.Npost[src], groupConfigs[0][srcGrp].numPostSynapses);
//      exitSimulation(1);
//  }
//
//  if(managerRuntimeData.Npre[dest] >= groupConfigs[0][destGrp].numPreSynapses) {
//      KERNEL_ERROR("setConnection(%d (Grp=%s), %d (Grp=%s), %f, %d)", src, groupInfo[srcGrp].Name.c_str(),
//                  dest, groupInfo[destGrp].Name.c_str(), synWt, dVal);
//      KERNEL_ERROR("Large number of presynaptic connections established (%d), max for this group %d.", managerRuntimeData.Npre[dest], groupConfigs[0][destGrp].numPreSynapses);
//      exitSimulation(1);
//  }
//
//  int p = managerRuntimeData.Npost[src];
//
//  assert(managerRuntimeData.Npost[src] >= 0);
//  assert(managerRuntimeData.Npre[dest] >= 0);
//  assert((src * maxNumPostSynGrp + p) / numN < maxNumPostSynGrp); // divide by numN to prevent INT overflow
//
//  unsigned int post_pos = managerRuntimeData.cumulativePost[src] + managerRuntimeData.Npost[src];
//  unsigned int pre_pos  = managerRuntimeData.cumulativePre[dest] + managerRuntimeData.Npre[dest];
//
//  assert(post_pos < numPostSynNet);
//  assert(pre_pos  < numPreSynNet);
//
//  //generate a new postSynapticIds id for the current connection
//  managerRuntimeData.postSynapticIds[post_pos]   = SET_CONN_ID(dest, managerRuntimeData.Npre[dest], destGrp);
//  tmp_SynapticDelay[post_pos] = dVal;
//
//  managerRuntimeData.preSynapticIds[pre_pos] = SET_CONN_ID(src, managerRuntimeData.Npost[src], srcGrp);
//  managerRuntimeData.wt[pre_pos]    = synWt;
//  managerRuntimeData.maxSynWt[pre_pos] = maxWt;
//  managerRuntimeData.connIdsPreIdx[pre_pos] = connId;
//
//  bool synWtType = GET_FIXED_PLASTIC(connProp);
//
//  if (synWtType == SYN_PLASTIC) {
//      sim_with_fixedwts = false; // if network has any plastic synapses at all, this will be set to true
//      managerRuntimeData.Npre_plastic[dest]++;
//      // homeostasis
//      if (groupConfigs[0][destGrp].WithHomeostasis && groupConfigs[0][destGrp].homeoId ==-1)
//          groupConfigs[0][destGrp].homeoId = dest; // this neuron info will be printed
//  }
//
//  managerRuntimeData.Npre[dest] += 1;
//  managerRuntimeData.Npost[src] += 1;
//
//  groupInfo[srcGrp].numPostConn++;
//  groupInfo[destGrp].numPreConn++;
//
//  if (managerRuntimeData.Npost[src] > groupInfo[srcGrp].maxPostConn)
//      groupInfo[srcGrp].maxPostConn = managerRuntimeData.Npost[src];
//  if (managerRuntimeData.Npre[dest] > groupInfo[destGrp].maxPreConn)
//  groupInfo[destGrp].maxPreConn = managerRuntimeData.Npre[src];
//}