void
ComputeMgr::createComputes(ComputeMap *map)
{
Node *node = Node::Object();
SimParameters *simParams = node->simParameters;
int myNode = node->myid();
ComputeNonbondedUtil::select();
if ( simParams->globalForcesOn && !myNode )
{
DebugM(4,"Mgr running on Node "<<CkMyPe()<<"\n");
/* create a master server to allow multiple masters */
masterServerObject = new GlobalMasterServer(this,
PatchMap::Object()->numNodesWithPatches());
/* create the individual global masters */
// masterServerObject->addClient(new GlobalMasterTest());
if (simParams->tclForcesOn)
masterServerObject->addClient(new GlobalMasterTcl());
if (simParams->IMDon && ! simParams->IMDignore)
masterServerObject->addClient(new GlobalMasterIMD());
if (simParams->SMDOn)
masterServerObject->addClient(
new GlobalMasterSMD(simParams->SMDk, simParams->SMDk2,
simParams->SMDVel,
simParams->SMDDir, simParams->SMDOutputFreq,
simParams->firstTimestep, simParams->SMDFile,
node->molecule->numAtoms)
);
if (simParams->symmetryOn &&
(simParams->firstTimestep < simParams->symmetryLastStep ||
simParams->symmetryLastStep == -1))
masterServerObject->addClient(new GlobalMasterSymmetry());
if (simParams->TMDOn)
masterServerObject->addClient(new GlobalMasterTMD());
if (simParams->miscForcesOn)
masterServerObject->addClient(new GlobalMasterMisc());
if ( simParams->freeEnergyOn )
masterServerObject->addClient(new GlobalMasterFreeEnergy());
if ( simParams->colvarsOn )
masterServerObject->addClient(new GlobalMasterColvars());
}
if ( !myNode && simParams->IMDon && simParams->IMDignore ) {
// GlobalMasterIMD constructor saves pointer to node->IMDOutput object
new GlobalMasterIMD();
}
#ifdef NAMD_CUDA
bool deviceIsMine = ( cuda_device_pe() == CkMyPe() );
#endif
// If there is a MIC device, then loop through the computes, flagging each one
// to be excuted on either a host or on a device
#ifdef NAMD_MIC
bool deviceIsMine = ( mic_device_pe() == CkMyPe() );
/*
#if MIC_SPLIT_WITH_HOST != 0
PatchMap *patchMap = PatchMap::Object();
int deviceThreshold = (patchMap->numaway_a() + patchMap->numaway_b() + patchMap->numaway_c()) - 2;
for (int i = 0; i < map->nComputes; i++) {
switch (map->type(i)) {
case computeNonbondedSelfType:
// Direct all non-bonded self computes to the device
map->setDirectToDevice(i, 1);
break;
case computeNonbondedPairType:
// Direct some non-bonded pair computes to the device, based on the manhattan distance
{
int pid0 = map->pid(i, 0);
int pid1 = map->pid(i, 1);
int index_a0 = patchMap->index_a(pid0);
int index_b0 = patchMap->index_b(pid0);
int index_c0 = patchMap->index_c(pid0);
int index_a1 = patchMap->index_a(pid1);
int index_b1 = patchMap->index_b(pid1);
int index_c1 = patchMap->index_c(pid1);
int da = index_a0 - index_a1; da *= ((da < 0) ? (-1) : (1));
int db = index_b0 - index_b1; db *= ((db < 0) ? (-1) : (1));
int dc = index_c0 - index_c1; dc *= ((dc < 0) ? (-1) : (1));
int manDist = da + db + dc;
map->setDirectToDevice(i, ((manDist <= deviceThreshold) ? (1) : (0)));
}
break;
default:
// All other computes should be directed to the host (flag is ignored, but set it)
map->setDirectToDevice(i, 0);
break;
} // end switch (map->type(i))
} // end for (i < map->nComputes)
#endif // MIC_SPLIT_WITH_HOST != 0
*/
#endif // NAMD_MIC
for (int i=0; i < map->nComputes; i++)
{
if ( ! ( i % 100 ) )
{
}
#if defined(NAMD_CUDA) || defined(NAMD_MIC)
switch ( map->type(i) )
{
#ifdef NAMD_CUDA
case computeNonbondedSelfType:
case computeNonbondedPairType:
if ( ! deviceIsMine ) continue;
if ( ! cuda_device_shared_with_pe(map->computeData[i].node) ) continue;
break;
#endif
#ifdef NAMD_MIC
case computeNonbondedSelfType:
#if MIC_SPLIT_WITH_HOST != 0
if (map->directToDevice(i)) { // If should be directed to the device...
#endif
if ( ! deviceIsMine ) continue;
if ( ! mic_device_shared_with_pe(map->computeData[i].node) ) continue;
#if MIC_SPLIT_WITH_HOST != 0
} else { // ... otherwise, direct to host...
if (map->computeData[i].node != myNode) { continue; }
}
#endif
break;
case computeNonbondedPairType:
#if MIC_SPLIT_WITH_HOST != 0
if (map->directToDevice(i)) { // If should be directed to the device...
#endif
if ( ! deviceIsMine ) continue;
if ( ! mic_device_shared_with_pe(map->computeData[i].node) ) continue;
#if MIC_SPLIT_WITH_HOST != 0
} else { // ... otherwise, direct to host...
if (map->computeData[i].node != myNode) { continue; }
}
#endif
break;
#endif
case computeNonbondedCUDAType:
case computeNonbondedMICType:
if ( ! deviceIsMine ) continue;
default:
if ( map->computeData[i].node != myNode ) continue;
}
#else
if ( map->computeData[i].node != myNode ) continue;
#endif
DebugM(1,"Compute " << i << '\n');
DebugM(1," node = " << map->computeData[i].node << '\n');
DebugM(1," type = " << map->computeData[i].type << '\n');
DebugM(1," numPids = " << map->computeData[i].numPids << '\n');
DebugM(1," numPidsAllocated = " << map->computeData[i].numPidsAllocated << '\n');
for (int j=0; j < map->computeData[i].numPids; j++)
{
// DebugM(1," pid " << map->computeData[i].pids[j] << '\n');
if (!((j+1) % 6))
DebugM(1,'\n');
}
DebugM(1,"\n---------------------------------------");
DebugM(1,"---------------------------------------\n");
createCompute(i, map);
}
#ifdef NAMD_CUDA
if ( computeNonbondedCUDAObject ) {
computeNonbondedCUDAObject->assignPatches();
}
#endif
#ifdef NAMD_MIC
if ( computeNonbondedMICObject ) {
computeNonbondedMICObject->assignPatches();
}
#endif
}
void
ComputeMgr::createComputes(ComputeMap *map)
{
Node *node = Node::Object();
SimParameters *simParams = node->simParameters;
int myNode = node->myid();
ComputeNonbondedUtil::select();
if ( simParams->globalForcesOn && !myNode )
{
DebugM(4,"Mgr running on Node "<<CkMyPe()<<"\n");
/* create a master server to allow multiple masters */
masterServerObject = new GlobalMasterServer(this,
PatchMap::Object()->numNodesWithPatches());
/* create the individual global masters */
// masterServerObject->addClient(new GlobalMasterTest());
if (simParams->tclForcesOn)
masterServerObject->addClient(new GlobalMasterTcl());
if (simParams->IMDon && ! simParams->IMDignore)
masterServerObject->addClient(new GlobalMasterIMD());
if (simParams->SMDOn)
masterServerObject->addClient(
new GlobalMasterSMD(simParams->SMDk, simParams->SMDk2,
simParams->SMDVel,
simParams->SMDDir, simParams->SMDOutputFreq,
simParams->firstTimestep, simParams->SMDFile,
node->molecule->numAtoms)
);
if (simParams->symmetryOn &&
(simParams->firstTimestep < simParams->symmetryLastStep ||
simParams->symmetryLastStep == -1))
masterServerObject->addClient(new GlobalMasterSymmetry());
if (simParams->TMDOn)
masterServerObject->addClient(new GlobalMasterTMD());
if (simParams->miscForcesOn)
masterServerObject->addClient(new GlobalMasterMisc());
if ( simParams->freeEnergyOn )
masterServerObject->addClient(new GlobalMasterFreeEnergy());
if ( simParams->colvarsOn )
masterServerObject->addClient(new GlobalMasterColvars());
}
if ( !myNode && simParams->IMDon && simParams->IMDignore ) {
// GlobalMasterIMD constructor saves pointer to node->IMDOutput object
new GlobalMasterIMD();
}
#ifdef NAMD_CUDA
bool deviceIsMine = ( cuda_device_pe() == CkMyPe() );
#endif
for (int i=0; i < map->nComputes; i++)
{
if ( ! ( i % 100 ) )
{
}
#ifdef NAMD_CUDA
switch ( map->type(i) )
{
case computeNonbondedSelfType:
case computeNonbondedPairType:
if ( ! deviceIsMine ) continue;
if ( ! cuda_device_shared_with_pe(map->computeData[i].node) ) continue;
break;
case computeNonbondedCUDAType:
if ( ! deviceIsMine ) continue;
default:
if ( map->computeData[i].node != myNode ) continue;
}
#else
if ( map->computeData[i].node != myNode ) continue;
#endif
DebugM(1,"Compute " << i << '\n');
DebugM(1," node = " << map->computeData[i].node << '\n');
DebugM(1," type = " << map->computeData[i].type << '\n');
DebugM(1," numPids = " << map->computeData[i].numPids << '\n');
DebugM(1," numPidsAllocated = " << map->computeData[i].numPidsAllocated << '\n');
for (int j=0; j < map->computeData[i].numPids; j++)
{
// DebugM(1," pid " << map->computeData[i].pids[j] << '\n');
if (!((j+1) % 6))
DebugM(1,'\n');
}
DebugM(1,"\n---------------------------------------");
DebugM(1,"---------------------------------------\n");
createCompute(i, map);
}
#ifdef NAMD_CUDA
if ( computeNonbondedCUDAObject ) {
computeNonbondedCUDAObject->assignPatches();
}
#endif
}