Revision 5dfe48ba8684b34080d95fa045be4282b8bdab08 authored by Harshitha on 16 March 2016, 19:50:35 UTC, committed by Harshitha on 16 March 2016, 19:50:35 UTC
Change-Id: I78ccaedcf3eea7e07b820ef4ae9ad8356ce11f88
1 parent b79547b
smooth.C
/*
* Implementation of the "smooth" operations needed for SPH.
*
* "smooth" itself is a search for the kth nearest neighbors. No
* calculations can be done until all the neighbors are found since
* the operations will depend on the distance to the kth nearest neighbor.
*
* Proposed strategy for bucketwise smooth (this is different than
* PKDGRAV, which does a particle at a time): each particle in a bucket
* maintains its own priority queue. Disadvantage: each "openCriterion"
* results in nBucket tests, but most of those tests have to be performed
* anyway.
*
* Priority queues are kept in the NearNeighborState class.
*
* How do I keep track of which particles are already in the queues?
*
* Discussion of strategies:
*
* PKDGRAV: primes the priority queue with particles in tree order
* starting with the bucket particles.
* Once a particle is finished, the nearest unfinished particle in the
* priority queue is done next ("the snake").
* This requires: 1) a "smooth done" bit for each particle. 2) This
* also requires a hash table to know if a particle is already in the
* queue.
*
* We would like to search more than one particle at once.
* That last bit means we need a hash table per particle, which seems
* overkill. We could avoid this by using a walk that 1) goes through
* the particles in order, and 2) skips the preloaded particles.
* Another way to avoid the hash table is to start the particle
* priority queue with a single particle that is known not to be in
* the final list, e.g. the 33rd closest particle. This is what is
* done below.
*/
#include "ParallelGravity.h"
#include "TreeWalk.h"
#include "DataManager.h"
#include "Opt.h"
#include "smooth.h"
#include "Space.h"
#include <float.h>
/// Point to the smooth parameters during a smooth walk. This is for
/// use by the EntryTypeSmoothParticle::unpack() method.
SmoothParams *globalSmoothParams;
/*
* There is actually not much "work" for the smooth walk. This method
* makes the decision about whether to keep a node or not. If we keep
* a node, and its a bucket, then we check each particle with
* bucketCompare().
*/
int SmoothCompute::doWork(GenericTreeNode *node, // Node to test
TreeWalk *tw,
State *state,
int chunk,
int reqID, // bucket we are working with
bool isRoot,
bool &didcomp, int awi)
{
// so that we have a quick return in case of empty nodes
if(node->getType() == Empty || node->getType() == CachedEmpty){
return DUMP;
}
if(!(node->iParticleTypes & params->iType)) {
return DUMP; // no particles of appropriate type
}
// check opening criterion
int open = openCriterion(tp, node, reqID, state);
// Turn open into an action
int action = opt->action(open, node);
if(action == KEEP_LOCAL_BUCKET && node->iRank != CkMyRank())
action = KEEP_REMOTE_BUCKET; // treat other rank's buckets as remote.
if(action == KEEP){
// Bounds intersect ball; descend to children
return KEEP;
}
else if(action == COMPUTE) {
CkAssert(0);
}
else if(action == KEEP_LOCAL_BUCKET) {
// Search bucket for contained particles
GravityParticle *part = node->particlePointer;
for(int i = node->firstParticle; i <= node->lastParticle; i++) {
if(!TYPETest(&part[i-node->firstParticle], params->iType))
continue;
// particle is part of search tree.
bucketCompare(tp, &part[i-node->firstParticle],
(GenericTreeNode *) computeEntity,
tp->getParticles(),
tp->decodeOffset(reqID),
state);
}
return DUMP;
}
else if(action == KEEP_REMOTE_BUCKET) {
GravityParticle *part;
part = tp->requestSmoothParticles(node->getKey(),
chunk,
node->remoteIndex,
node->firstParticle,
node->lastParticle,
reqID, awi, computeEntity, false);
if(part) {
// Particles available; Search for contained.
for(int i = node->firstParticle; i <= node->lastParticle; i++) {
if(!TYPETest(&part[i-node->firstParticle], params->iType))
continue;
// particle is part of search tree.
bucketCompare(tp, &part[i-node->firstParticle],
(GenericTreeNode *) computeEntity,
tp->getParticles(),
tp->decodeOffset(reqID),
state);
}
}
else {
// Missed cache; record this
state->counterArrays[0][decodeReqID(reqID)]
+= node->lastParticle - node->firstParticle+1;
}
return DUMP;
}
else if(action == DUMP) {
return DUMP;
}
CkAbort("SmoothCompute: bad walk");
return -1;
}
/*
* reassociate bucket
*/
void SmoothCompute::reassoc(void *ce, int ar, Opt *o){
computeEntity = ce;
activeRung = ar;
opt = o;
}
void SmoothCompute::nodeMissedEvent(int reqID, int chunk, State *state, TreePiece *tp)
{
int reqIDlist = decodeReqID(reqID);
state->counterArrays[0][reqIDlist]++;
}
// called after constructor, so tp should be set
State *KNearestSmoothCompute::getNewState(int nBuckets){
NearNeighborState *state = new NearNeighborState(tp->myNumParticles+2, nSmooth);
// array to keep track of outstanding requests
state->counterArrays[0] = new int [nBuckets];
state->counterArrays[1] = 0;
for(int j = 0; j < nBuckets; ++j){
state->counterArrays[0][j] = 1; // so we know that the local
// walk is not finished.
}
state->started = true;
CkAssert(tp->myTreeParticles >= 0);
state->nParticlesPending = tp->myTreeParticles;
state->currentBucket = 0;
state->bWalkDonePending = 0;
return state;
}
/*
* Return true if a sphere centered at pos with radius^2 of rsq
* intersects the bounding box of node.
* This is being used instead of the Space intersect method so we can use
* radius^2 for efficiency.
*/
static inline bool
intersect(OrientedBox<double>& box, Vector3D<double> pos, double rsq)
{
double dsq = 0.0;
double delta;
if((delta = box.lesser_corner.x - pos.x) > 0)
dsq += delta * delta;
else if((delta = pos.x - box.greater_corner.x) > 0)
dsq += delta * delta;
if(rsq < dsq)
return false;
if((delta = box.lesser_corner.y - pos.y) > 0)
dsq += delta * delta;
else if((delta = pos.y - box.greater_corner.y) > 0)
dsq += delta * delta;
if(rsq < dsq)
return false;
if((delta = box.lesser_corner.z - pos.z) > 0)
dsq += delta * delta;
else if((delta = pos.z - box.greater_corner.z) > 0)
dsq += delta * delta;
return (dsq <= rsq);
}
/**
* Opening criterion for the smoothBucket walk.
* Return true if we must open the node.
*/
int KNearestSmoothCompute::openCriterion(TreePiece *ownerTP,
GenericTreeNode *node, // Node to test
int reqID, State *state) {
GenericTreeNode *myNode = (GenericTreeNode *) computeEntity;
GravityParticle *particles = ownerTP->getParticles();
Vector3D<double> offset = ownerTP->decodeOffset(reqID);
NearNeighborState *nstate = (NearNeighborState *)state;
double rBucket = myNode->sizeSm + myNode->fKeyMax;
if(!intersect(node->boundingBox, myNode->centerSm - offset,
rBucket*rBucket))
return 0;
for(int j = myNode->firstParticle; j <= myNode->lastParticle; ++j) {
if(!params->isSmoothActive(&particles[j]))
continue;
double r2 = nstate->Qs[j][0].fKey; // Ball radius^2
if(intersect(node->boundingBox, particles[j].position - offset, r2)) {
return 1;
}
}
return 0;
}
inline double sqr(double x) { return x*x; }
/**
* Test a given particle against all the priority queues in the
* bucket.
*/
void KNearestSmoothCompute::bucketCompare(TreePiece *ownerTP,
GravityParticle *p, // Particle to test
GenericTreeNode *node, // bucket
GravityParticle *particles, // local
// particle data
Vector3D<double> offset,
State *state
)
{
NearNeighborState *nstate = (NearNeighborState *)state;
Vector3D<double> rp = offset + p->position;
Vector3D<double> drBucket = node->centerSm - rp;
if(sqr(node->sizeSm + node->fKeyMax) < drBucket.lengthSquared())
return; // particle is outside all smoothing radii
double dKeyMaxBucket = 0.0;
for(int j = node->firstParticle; j <= node->lastParticle; ++j) {
if(!params->isSmoothActive(&particles[j]))
continue;
CkVec<pqSmoothNode> &Q = nstate->Qs[j];
double rOld2 = Q[0].fKey; // Ball radius
Vector3D<double> dr = particles[j].position - rp;
// include particle if less than the current search radius, or
// less than the h_min limit set by softening.
if(rOld2 >= dr.lengthSquared()) {
pqSmoothNode pqNew;
pqNew.fKey = dr.lengthSquared();
pqNew.dx = dr;
pqNew.p = p;
// Perform replacement if we've got enough particles and
// we are not hitting the h_min limit.
if(iLowhFix && rOld2 <= dfBall2OverSoft2*sqr(particles[j].soft)) {
Q.push_back(pqNew);
Q[0].fKey = dfBall2OverSoft2*sqr(particles[j].soft);
}
else {
if(Q.size() >= nSmooth) {
std::pop_heap(&(Q[0]) + 0, &(Q[0]) + nSmooth);
Q.resize(Q.size()-1); // pop if list is full
}
Q.push_back(pqNew);
std::push_heap(&(Q[0]) + 0, &(Q[0]) + Q.size());
}
}
if(Q[0].fKey > dKeyMaxBucket)
dKeyMaxBucket = Q[0].fKey;
}
node->fKeyMax = sqrt(dKeyMaxBucket);
}
/**
* Process particles received from missed Cache request
*/
void KNearestSmoothCompute::recvdParticlesFull(GravityParticle *part,
int num, int chunk,int reqID, State *state,
TreePiece *tp, Tree::NodeKey &remoteBucket){
Vector3D<double> offset = tp->decodeOffset(reqID);
int reqIDlist = decodeReqID(reqID);
CkAssert(num > 0);
state->counterArrays[0][reqIDlist] -= num;
GenericTreeNode* reqnode = tp->bucketList[reqIDlist];
for(int i=0;i<num;i++){
if(!TYPETest(&part[i], params->iType))
continue;
bucketCompare(tp, &part[i], reqnode, tp->myParticles, offset, state);
}
((NearNeighborState *)state)->finishBucketSmooth(reqIDlist, tp);
}
void KNearestSmoothCompute::nodeRecvdEvent(TreePiece *owner, int chunk, State *state,
int reqIDlist){
state->counterArrays[0][reqIDlist]--;
((NearNeighborState *)state)->finishBucketSmooth(reqIDlist, owner);
}
void TreePiece::setupSmooth() {
// XXX I don't believe any of the Chunks are used in the smooth walk.
dm->getChunks(numChunks, prefetchRoots);
CkArrayIndexMax idxMax = CkArrayIndex1D(thisIndex);
if (numChunks == 0 && myNumParticles == 0) numChunks = 1;
cacheSmoothPart.ckLocalBranch()->cacheSync(numChunks, idxMax, localIndex);
cacheNode.ckLocalBranch()->cacheSync(numChunks, idxMax, localIndex);
// The following if is necessary to prevent nodes containing only TreePieces
// without particles from getting stuck and crashing.
if (myNumParticles == 0) {
// No particles assigned to this TreePiece
for (int i=0; i< numChunks; ++i) {
cacheSmoothPart[CkMyPe()].finishedChunk(i, 0);
}
return;
}
}
// Start tree walk and smooth calculation
void TreePiece::startSmooth(// type of smoothing and parameters
SmoothParams* params,
int iLowhFix,
int nSmooth,
double dfBall2OverSoft2,
const CkCallback& cb) {
LBTurnInstrumentOn(); // Be sure the Load Balancer is running.
CkAssert(nSmooth > 0);
cbSmooth = cb;
activeRung = params->activeRung;
try {
setupSmooth();
if (myNumParticles == 0) {
markSmoothWalkDone();
return;
}
// Create objects that are reused by all buckets
twSmooth = new BottomUpTreeWalk;
sSmooth = new KNearestSmoothCompute(this, params, nSmooth, iLowhFix,
dfBall2OverSoft2);
initBucketsSmooth(sSmooth);
// creates and initializes nearneighborstate object
sSmoothState = sSmooth->getNewState(numBuckets);
optSmooth = new SmoothOpt;
processReqSmoothParticles();
activeWalks.reserve(maxAwi);
addActiveWalk(smoothAwi, twSmooth,sSmooth,optSmooth,sSmoothState);
#ifdef CHECK_WALK_COMPLETIONS
CkPrintf("[%d] addActiveWalk smooth (%d)\n", thisIndex, activeWalks.length());
#endif
calculateSmoothLocal();
}
catch (std::bad_alloc) {
CkAbort("Out of memory in startSmooth");
}
}
// Initialize particles in bucket and bucket bookkeeping
// Template for both SmoothCompute and ReSmoothCompute types.
template <class Tsmooth>
void TreePiece::initBucketsSmooth(Tsmooth tSmooth) {
tSmooth->nActive = 0;
for (unsigned int j=0; j<numBuckets; ++j) {
GenericTreeNode* node = bucketList[j];
// TODO: active bounds may give a performance boost in the
// multi-timstep regime.
// node->boundingBox.reset(); // XXXX dangerous should have separate
// Active bounds
for(int i = node->firstParticle; i <= node->lastParticle; ++i) {
if(tSmooth->params->isSmoothActive(&myParticles[i])) {
tSmooth->nActive++;
tSmooth->params->initSmoothParticle(&myParticles[i]);
// node->boundingBox.grow(myParticles[i].position);
}
else if (TYPETest(&myParticles[i], tSmooth->params->iType)) {
tSmooth->params->initTreeParticle(&myParticles[i]);
}
}
}
}
// Start the smoothing
void TreePiece::calculateSmoothLocal() {
dummyMsg *msg = new (8*sizeof(int)) dummyMsg;
// Give smooths higher priority than gravity
*((int *)CkPriorityPtr(msg)) = thisIndex + 1;
CkSetQueueing(msg,CK_QUEUEING_IFIFO);
msg->val=0;
thisProxy[thisIndex].nextBucketSmooth(msg);
}
//
// Do the next set of buckets
//
void TreePiece::nextBucketSmooth(dummyMsg *msg){
unsigned int i=0;
int currentBucket = sSmoothState->currentBucket;
// smooths are faster than gravity, and they need cache messages to
// get through. Therefore yield after every bucket.
while(i<1 && currentBucket<numBuckets){
smoothNextBucket();
currentBucket++;
sSmoothState->currentBucket++;
i++;
}
if (currentBucket<numBuckets) { // Queue up the next set
thisProxy[thisIndex].nextBucketSmooth(msg);
} else {
delete msg;
}
}
void KNearestSmoothCompute::initSmoothPrioQueue(int iBucket, State *state)
{
// Prime the queues
GenericTreeNode *myNode = tp->bucketList[iBucket];
// state is passed in to function now.
NearNeighborState *nstate = (NearNeighborState *)state;
int bucketActive = 0;
for(int j = myNode->firstParticle; j <= myNode->lastParticle; ++j) {
if(params->isSmoothActive(&tp->myParticles[j]))
bucketActive++;
}
if(!bucketActive) // No particles in this bucket are active.
return;
//
// Get nearest nSmooth particles in tree order
//
int firstQueue = myNode->firstParticle;
int iCount = 0;
int lastQueue = firstQueue;
// Find only particles of interest
// First search from the start of the bucket to the end of this treepiece
// N.B. We are searching for nSmooth+1 particles.
for(lastQueue = firstQueue;
iCount <= nSmooth && lastQueue <= tp->myNumParticles;
lastQueue++) {
if(TYPETest(&tp->myParticles[lastQueue], params->iType))
iCount++;
}
// We still don't have enough particles. Search back to the start of the
// treepiece
int bEnough = 1; // Do we have enough particles on piece to get a limit?
if(lastQueue > tp->myNumParticles)
{
firstQueue = myNode->firstParticle;
// N.B. We are searching for nSmooth+1 particles.
while(iCount <= nSmooth) {
firstQueue--;
if(firstQueue == 0) {
bEnough = 0; // Ran out of particles
firstQueue++;
break;
}
if(TYPETest(&tp->myParticles[firstQueue], params->iType))
iCount++;
}
}
if(bEnough && ((lastQueue - firstQueue) <= nSmooth))
CkAbort("Missing particles");
CkAssert(firstQueue > 0);
CkAssert(lastQueue <= tp->myNumParticles+1);
OrientedBox<double> bndSmoothAct; // bounding box for smoothActive particles
double dKeyMaxBucket = 0.0;
for(int j = myNode->firstParticle; j <= myNode->lastParticle; ++j) {
if(!params->isSmoothActive(&tp->myParticles[j]))
continue;
GravityParticle *p = &tp->myParticles[j];
bndSmoothAct.grow(p->position);
CkVec<pqSmoothNode> *Q = &nstate->Qs[j];
Q->reserve(nSmooth);
//
// Find maximum of nearest neighbors
//
double drMax2 = 0.0;
for(int k = firstQueue; k < lastQueue; k++)
{
if(!TYPETest(&tp->myParticles[k], params->iType))
continue;
Vector3D<double> dr = tp->myParticles[k].position - p->position;
if(dr.lengthSquared() > drMax2) {
drMax2 = dr.lengthSquared();
}
}
CkAssert(Q->size() == 0);
pqSmoothNode pqNew;
if(bEnough)
pqNew.fKey = drMax2;
else
pqNew.fKey = DBL_MAX;
//
// For FastGas, we have a limit on the size of the search ball.
//
if(params->bUseBallMax && p->isGas() && p->fBallMax() > 0.0
&& pqNew.fKey > p->fBallMax()*p->fBallMax())
pqNew.fKey = p->fBallMax()*p->fBallMax();
if(iLowhFix && pqNew.fKey < dfBall2OverSoft2*p->soft*p->soft)
pqNew.fKey = dfBall2OverSoft2*p->soft*p->soft;
if(pqNew.fKey > dKeyMaxBucket)
dKeyMaxBucket = pqNew.fKey;
pqNew.p = NULL;
Q->push_back(pqNew);
std::push_heap(&((*Q)[0]) + 0, &((*Q)[0]) + 1);
}
myNode->centerSm = bndSmoothAct.center();
myNode->sizeSm = .5*(bndSmoothAct.size()).length();
myNode->fKeyMax = sqrt(dKeyMaxBucket);
}
void TreePiece::smoothBucketComputation() {
twSmooth->init(sSmooth, this);
int currentBucket = sSmoothState->currentBucket;
GenericTreeNode *myNode = bucketList[currentBucket];
sSmooth->init(myNode, activeRung, optSmooth);
int bucketActive = 0;
for(int j = myNode->firstParticle; j <= myNode->lastParticle; ++j) {
if(sSmooth->params->isSmoothActive(&myParticles[j]))
bucketActive += 1;
}
// start the tree walk from the tree built in the cache
if (bucketActive) {
for(int cr = 0; cr < numChunks; cr++){
GenericTreeNode *chunkRoot = dm->chunkRootToNode(prefetchRoots[cr]);
if(!chunkRoot){
continue;
}
twSmooth->walk(chunkRoot, sSmoothState, cr,
encodeOffset(currentBucket, 0,0,0), smoothAwi);
for(int x = -nReplicas; x <= nReplicas; x++) {
for(int y = -nReplicas; y <= nReplicas; y++) {
for(int z = -nReplicas; z <= nReplicas; z++) {
if(x || y || z)
twSmooth->walk(chunkRoot, sSmoothState, cr,
encodeOffset(currentBucket, x,y,z), smoothAwi);
}
}
}
}
}
sSmoothState->counterArrays[0][currentBucket]--;
}
void TreePiece::smoothNextBucket() {
int currentBucket = sSmoothState->currentBucket;
if(currentBucket >= numBuckets)
return;
((KNearestSmoothCompute *)sSmooth)->initSmoothPrioQueue(currentBucket,
sSmoothState);
smoothBucketComputation();
((NearNeighborState *)sSmoothState)->finishBucketSmooth(currentBucket, this);
}
void NearNeighborState::finishBucketSmooth(int iBucket, TreePiece *tp) {
GenericTreeNode *node = tp->bucketList[iBucket];
if(counterArrays[0][iBucket] == 0) {
tp->sSmooth->walkDone(this);
if(verbosity>4)
CkPrintf("[%d] TreePiece %d finished smooth with bucket %d\n",CkMyPe(),
tp->thisIndex,iBucket);
nParticlesPending -= node->particleCount;
if(started && nParticlesPending == 0) {
started = false;
#ifdef CACHE_MEM_STATS
tp->memWithCache = CmiMemoryUsage()/(1024*1024);
#endif
tp->nNodeCacheEntries = cacheNode.ckLocalBranch()->getCache()->size();
tp->nPartCacheEntries = cacheSmoothPart.ckLocalBranch()->getCache()->size();
cacheSmoothPart[CkMyPe()].finishedChunk(0, 0);
#ifdef CHECK_WALK_COMPLETIONS
CkPrintf("[%d] markWalkDone NearNeighborState\n", tp->getIndex());
#endif
tp->markSmoothWalkDone();
if(verbosity > 1)
ckerr << "TreePiece " << tp->thisIndex << ": My particles are done"
<< endl;
}
}
}
void TreePiece::markSmoothWalkDone()
{
CkCallback cb = CkCallback(CkIndex_TreePiece::finishSmoothWalk(),
pieces);
// Use shadow array to avoid reduction conflict
smoothProxy[thisIndex].ckLocal()->contribute(cb);
}
void TreePiece::finishSmoothWalk()
{
// Furthermore, we need to wait for outstanding flushes to be
// processed for the combiner cache.
if(sSmooth && nCacheAccesses > 0) {
sSmoothState->bWalkDonePending = 1;
return;
}
// At this point, this piece has finished its walk, and all particle
// requests from other processors have been flushed back through the
// combiner.
for(int i = 0; i < myNumParticles; i++) {
sSmooth->params->postTreeParticle(&myParticles[i+1]);
}
#ifdef CACHE_MEM_STATS
memPostCache = CmiMemoryUsage()/(1024*1024);
#endif
nCacheAccesses = 0; // reset for next walk.
if(myNumParticles != 0) {
sSmooth->freeState(sSmoothState);
delete sSmooth;
delete optSmooth;
delete twSmooth;
sSmooth = 0;
}
#ifdef CHECK_WALK_COMPLETIONS
CkPrintf("[%d] inside finishSmoothWalk contrib callback\n", thisIndex);
#endif
smoothProxy[thisIndex].ckLocal()->contribute(cbSmooth);
}
void KNearestSmoothCompute::walkDone(State *state) {
GenericTreeNode *node = (GenericTreeNode *) computeEntity;
GravityParticle *part = node->particlePointer;
for(int i = node->firstParticle; i <= node->lastParticle; i++) {
GravityParticle *p = &part[i-node->firstParticle];
if(!params->isSmoothActive(p))
continue;
NearNeighborState *nstate = (NearNeighborState *)state;
CkVec<pqSmoothNode> &Q = nstate->Qs[i];
double h = sqrt(Q[0].fKey); // Ball radius
int nCnt = Q.size();
if(Q[0].p == NULL) { // This can happen if iLowhFix is
// operating
CkAssert(iLowhFix != 0);
std::pop_heap(&(Q[0]) + 0, &(Q[0]) + nCnt);
nCnt--;
}
// The following assert() is not valid with the FastGas optimization.
// CkAssert(nCnt >= nSmooth);
// Limit fBall growth to help stability and inverse neighbor finding.
if(!(iLowhFix && h*h <= dfBall2OverSoft2*p->soft*p->soft)
&& params->bUseBallMax && p->isGas() && p->fBallMax() > 0.0
&& h > p->fBallMax()) {
h = p->fBallMax();
// With round-off, we might be in the LowhFix regime and
// Q might not be in heap order.
std::make_heap(&(Q[0]) + 0, &(Q[0]) + nCnt);
while(Q[0].fKey > h*h) {
std::pop_heap(&(Q[0]) + 0, &(Q[0]) + nCnt);
nCnt--;
}
}
p->fBall = h;
params->fcnSmooth(p, nCnt, &(Q[0]));
Q.clear();
}
// XXX jetley - the nearneighborstate allocated for this compute
// may be deleted after this function is called.
}
// From here down are "ReSmooth" methods.
// called after constructor, so tp should be set
State *ReSmoothCompute::getNewState(int nBucket){
ReNearNeighborState *state = new ReNearNeighborState(tp->myNumParticles+2);
// array to keep track of outstanding requests
state->counterArrays[0] = new int [nBucket];
state->counterArrays[1] = 0;
for (int j = 0; j < nBucket; ++j) {
state->counterArrays[0][j] = 1; // so we know that the local
// walk is not finished.
}
state->started = true;
CkAssert(tp->myTreeParticles >= 0);
state->nParticlesPending = tp->myTreeParticles;
state->currentBucket = 0;
state->bWalkDonePending = 0;
return state;
}
/*
* Opening criterion for the reSmoothBucket walk.
* Return true if we must open the node.
*/
int ReSmoothCompute::openCriterion(TreePiece *ownerTP,
GenericTreeNode *node, // Node to test
int reqID, State *state) {
GenericTreeNode *myNode = (GenericTreeNode *) computeEntity;
GravityParticle *particles = ownerTP->getParticles();
Vector3D<double> offset = ownerTP->decodeOffset(reqID);
double rBucket = myNode->sizeSm + myNode->fKeyMax;
if(!intersect(node->boundingBox, myNode->centerSm - offset,
rBucket*rBucket))
return 0;
for(int j = myNode->firstParticle; j <= myNode->lastParticle; ++j) {
if(!params->isSmoothActive(&particles[j]))
continue;
double r = particles[j].fBall; // Ball radius
if(intersect(node->boundingBox, particles[j].position - offset, r*r)) {
return 1;
}
}
return 0;
}
/*
* Test a given particle against all the priority queues in the
* bucket.
*/
void ReSmoothCompute::bucketCompare(TreePiece *ownerTP,
GravityParticle *p, // Particle to test
GenericTreeNode *node, // bucket
GravityParticle *particles, // local
// particle data
Vector3D<double> offset,
State *state
)
{
const double dSearchEps = 1e-7; // Slop to insure farthest neighbor is
// found.
ReNearNeighborState *nstate = (ReNearNeighborState *)state;
Vector3D<double> rp = offset + p->position;
Vector3D<double> drBucket = node->centerSm - rp;
if(sqr(node->sizeSm + node->fKeyMax)*(1.+dSearchEps) < drBucket.lengthSquared())
return; // particle is outside all smoothing radii
for(int j = node->firstParticle; j <= node->lastParticle; ++j) {
if(!params->isSmoothActive(&particles[j]))
continue;
CkVec<pqSmoothNode> *Q = &nstate->Qs[j];
double rOld = particles[j].fBall; // Ball radius
Vector3D<double> dr = particles[j].position - rp;
if(rOld*rOld*(1.+dSearchEps) >= dr.lengthSquared()) { // Add to list
pqSmoothNode pqNew;
pqNew.fKey = dr.lengthSquared();
pqNew.dx = dr;
pqNew.p = p;
Q->push_back(pqNew);
}
}
}
/*
* Process particles received from missed Cache request
*/
void ReSmoothCompute::recvdParticlesFull(GravityParticle *part,
int num, int chunk,int reqID, State *state,
TreePiece *tp, Tree::NodeKey &remoteBucket){
Vector3D<double> offset = tp->decodeOffset(reqID);
int reqIDlist = decodeReqID(reqID);
CkAssert(num > 0);
state->counterArrays[0][reqIDlist] -= num;
GenericTreeNode* reqnode = tp->bucketList[reqIDlist];
for(int i=0;i<num;i++){
if(!TYPETest(&part[i], params->iType))
continue;
bucketCompare(tp, &part[i], reqnode, tp->myParticles, offset, state);
}
((ReNearNeighborState *)state)->finishBucketSmooth(reqIDlist, tp);
}
void ReSmoothCompute::nodeRecvdEvent(TreePiece *owner, int chunk, State *state,
int reqIDlist){
state->counterArrays[0][reqIDlist]--;
((ReNearNeighborState *)state)->finishBucketSmooth(reqIDlist, owner);
}
// Start tree walk and smooth calculation
void TreePiece::startReSmooth(SmoothParams* params,
const CkCallback& cb) {
cbSmooth = cb;
activeRung = params->activeRung;
setupSmooth();
if (myNumParticles == 0) {
markSmoothWalkDone();
return;
}
//for (int i=0; i<numChunks; ++i) remaining Chunk[i] = myNumParticles;
// Create objects that are reused by all buckets
twSmooth = new TopDownTreeWalk;
sSmooth = new ReSmoothCompute(this, params);
initBucketsSmooth(sSmooth);
// creates and initializes nearneighborstate object
sSmoothState = sSmooth->getNewState(numBuckets);
optSmooth = new SmoothOpt;
processReqSmoothParticles();
activeWalks.reserve(maxAwi);
addActiveWalk(smoothAwi, twSmooth,sSmooth,optSmooth,sSmoothState);
#ifdef CHECK_WALK_COMPLETIONS
CkPrintf("[%d] addActiveWalk reSmooth (%d)\n", thisIndex, activeWalks.length());
#endif
thisProxy[thisIndex].calculateReSmoothLocal();
}
// Start the smoothing
void TreePiece::calculateReSmoothLocal() {
dummyMsg *msg = new (8*sizeof(int)) dummyMsg;
// Give smooths higher priority than gravity
*((int *)CkPriorityPtr(msg)) = thisIndex + 1;
CkSetQueueing(msg,CK_QUEUEING_IFIFO);
msg->val=0;
thisProxy[thisIndex].nextBucketReSmooth(msg);
}
//
// Do the next set of buckets
//
void TreePiece::nextBucketReSmooth(dummyMsg *msg){
unsigned int i=0;
int currentBucket = sSmoothState->currentBucket;
while(i<_yieldPeriod && currentBucket<numBuckets){
reSmoothNextBucket();
currentBucket++;
sSmoothState->currentBucket++;
i++;
}
if (currentBucket<numBuckets) { // Queue up the next set
thisProxy[thisIndex].nextBucketReSmooth(msg);
} else {
delete msg;
}
}
void TreePiece::reSmoothNextBucket() {
int currentBucket = sSmoothState->currentBucket;
if(currentBucket >= numBuckets)
return;
// set bucket search quantities
GenericTreeNode *myNode = bucketList[currentBucket];
OrientedBox<double> bndSmoothAct; // bounding box for smoothActive particles
double dKeyMaxBucket = 0.0;
int bucketActive = 0;
for(int j = myNode->firstParticle; j <= myNode->lastParticle; ++j) {
if(!sSmooth->params->isSmoothActive(&myParticles[j]))
continue;
bucketActive++;
bndSmoothAct.grow(myParticles[j].position);
if(myParticles[j].fBall > dKeyMaxBucket)
dKeyMaxBucket = myParticles[j].fBall;
}
if (bucketActive != 0) {
myNode->centerSm = bndSmoothAct.center();
myNode->sizeSm = .5*(bndSmoothAct.size()).length();
myNode->fKeyMax = dKeyMaxBucket;
}
smoothBucketComputation();
((ReNearNeighborState *)sSmoothState)->finishBucketSmooth(currentBucket, this);
}
void ReNearNeighborState::finishBucketSmooth(int iBucket, TreePiece *tp) {
GenericTreeNode *node = tp->bucketList[iBucket];
if(counterArrays[0][iBucket] == 0) {
tp->sSmooth->walkDone(this);
nParticlesPending -= node->particleCount;
if(verbosity>4)
CkPrintf("[%d] TreePiece %d finished resmooth with bucket %d, %d Pending\n",CkMyPe(),
tp->thisIndex,iBucket,nParticlesPending);
if(started && nParticlesPending == 0) {
started = false;
#ifdef CACHE_MEM_STATS
tp->memWithCache = CmiMemoryUsage()/(1024*1024);
#endif
cacheSmoothPart[CkMyPe()].finishedChunk(0, 0);
#ifdef CHECK_WALK_COMPLETIONS
CkPrintf("[%d] markWalkDone ReNearNeighborState\n", tp->getIndex());
#endif
tp->markSmoothWalkDone();
if(verbosity > 1)
ckerr << "TreePiece " << tp->thisIndex << ": My particles are done"
<< endl;
}
}
}
void ReSmoothCompute::walkDone(State *state) {
GenericTreeNode *node = (GenericTreeNode *) computeEntity;
GravityParticle *part = node->particlePointer;
for(int i = node->firstParticle; i <= node->lastParticle; i++) {
if(!params->isSmoothActive(&part[i-node->firstParticle]))
continue;
CkVec<pqSmoothNode> *Q = &((ReNearNeighborState *)state)->Qs[i];
pqSmoothNode *NN = NULL;
int nCnt = Q->size();
if(nCnt > 0)
NN = &((*Q)[0]);
params->fcnSmooth(&part[i-node->firstParticle], nCnt, NN);
Q->clear();
}
}
// called after constructor, so tp should be set
State *MarkSmoothCompute::getNewState(int nBucket){
MarkNeighborState *state = new MarkNeighborState(tp->myNumParticles+2);
state->counterArrays[0] = new int [nBucket];
state->counterArrays[1] = 0;
for (int j = 0; j < nBucket; ++j) {
state->counterArrays[0][j] = 1; // so we know that the local
// walk is not finished.
}
state->started = true;
CkAssert(tp->myTreeParticles >= 0);
state->nParticlesPending = tp->myTreeParticles;
state->currentBucket = 0;
state->bWalkDonePending = 0;
return state;
}
/*
* Opening criterion for the MarkSmoothBucket walk.
* Return true if we must open the node.
*/
int MarkSmoothCompute::openCriterion(TreePiece *ownerTP,
GenericTreeNode *node, // Node to test
int reqID, State *state) {
GenericTreeNode *myNode = (GenericTreeNode *) computeEntity; // my bucket
GravityParticle *particles = ownerTP->getParticles();
Vector3D<double> offset = ownerTP->decodeOffset(reqID);
for(int j = myNode->firstParticle; j <= myNode->lastParticle; ++j) {
if(!params->isSmoothActive(&particles[j]))
continue;
Vector3D<double> p(particles[j].position - offset);
if(Space::contains(node->bndBoxBall, p)) {
return 1;
}
}
return 0;
}
/*
* Test a given particle against all the priority queues in the
* bucket.
*/
void MarkSmoothCompute::bucketCompare(TreePiece *ownerTP,
GravityParticle *p, // Particle to test
GenericTreeNode *node, // bucket
GravityParticle *particles, // local
// particle data
Vector3D<double> offset,
State *state
)
{
Vector3D<double> rp = offset + p->position;
for(int j = node->firstParticle; j <= node->lastParticle; ++j) {
if(!params->isSmoothActive(&particles[j]))
continue;
double rOld = p->fBallMax(); // Ball radius
Vector3D<double> dr = particles[j].position - rp;
if(rOld*rOld >= dr.lengthSquared()) { // Mark as Neighbor
TYPESet(p, TYPE_NbrOfACTIVE);
}
}
}
/*
* Process particles received from missed Cache request
*/
void MarkSmoothCompute::recvdParticlesFull(GravityParticle *part,
int num, int chunk,int reqID, State *state,
TreePiece *tp, Tree::NodeKey &remoteBucket){
Vector3D<double> offset = tp->decodeOffset(reqID);
int reqIDlist = decodeReqID(reqID);
CkAssert(num > 0);
state->counterArrays[0][reqIDlist] -= num;
GenericTreeNode* reqnode = tp->bucketList[reqIDlist];
for(int i=0;i<num;i++){
if(!TYPETest(&part[i], params->iType))
continue;
bucketCompare(tp, &part[i], reqnode, tp->myParticles, offset, state);
}
((MarkNeighborState *)state)->finishBucketSmooth(reqIDlist, tp);
}
void MarkSmoothCompute::nodeRecvdEvent(TreePiece *owner, int chunk,
State *state, int reqIDlist){
state->counterArrays[0][reqIDlist]--;
((MarkNeighborState *)state)->finishBucketSmooth(reqIDlist, owner);
}
// Start marksmooth walk
void TreePiece::startMarkSmooth(SmoothParams* params,
const CkCallback& cb) {
cbSmooth = cb;
activeRung = params->activeRung;
setupSmooth();
if (myNumParticles == 0) {
markSmoothWalkDone();
return;
}
//for (int i=0; i<numChunks; ++i) remaining Chunk[i] = myNumParticles;
// Create objects that are reused by all buckets
twSmooth = new TopDownTreeWalk;
sSmooth = new MarkSmoothCompute(this, params);
initBucketsSmooth(sSmooth);
// creates and initializes nearneighborstate object
sSmoothState = sSmooth->getNewState(numBuckets);
optSmooth = new SmoothOpt;
processReqSmoothParticles();
addActiveWalk(smoothAwi, twSmooth,sSmooth,optSmooth,sSmoothState);
#ifdef CHECK_WALK_COMPLETIONS
CkPrintf("[%d] addActiveWalk reSmooth (%d)\n", thisIndex, activeWalks.length());
#endif
thisProxy[thisIndex].calculateMarkSmoothLocal();
}
// Start the smoothing
void TreePiece::calculateMarkSmoothLocal() {
dummyMsg *msg = new (8*sizeof(int)) dummyMsg;
// Give smooths higher priority than gravity
*((int *)CkPriorityPtr(msg)) = thisIndex + 1;
CkSetQueueing(msg,CK_QUEUEING_IFIFO);
msg->val=0;
thisProxy[thisIndex].nextBucketMarkSmooth(msg);
}
//
// Do the next set of buckets
//
void TreePiece::nextBucketMarkSmooth(dummyMsg *msg){
unsigned int i=0;
int currentBucket = sSmoothState->currentBucket;
while(i<_yieldPeriod && currentBucket<numBuckets){
markSmoothNextBucket();
currentBucket++;
sSmoothState->currentBucket++;
i++;
}
if (currentBucket<numBuckets) { // Queue up the next set
thisProxy[thisIndex].nextBucketMarkSmooth(msg);
} else {
delete msg;
}
}
void TreePiece::markSmoothNextBucket() {
int currentBucket = sSmoothState->currentBucket;
if(currentBucket >= numBuckets)
return;
smoothBucketComputation();
((MarkNeighborState *)sSmoothState)->finishBucketSmooth(currentBucket, this);
}
void MarkNeighborState::finishBucketSmooth(int iBucket, TreePiece *tp) {
GenericTreeNode *node = tp->bucketList[iBucket];
if(counterArrays[0][iBucket] == 0) {
tp->sSmooth->walkDone(this);
if(verbosity>4)
CkPrintf("[%d] TreePiece %d finished smooth with bucket %d\n",CkMyPe(),
tp->thisIndex,iBucket);
nParticlesPending -= node->particleCount;
if(started && nParticlesPending == 0) {
started = false;
#ifdef CACHE_MEM_STATS
tp->memWithCache = CmiMemoryUsage()/(1024*1024);
#endif
cacheSmoothPart[CkMyPe()].finishedChunk(0, 0);
#ifdef CHECK_WALK_COMPLETIONS
CkPrintf("[%d] markWalkDone ReNearNeighborState\n", tp->getIndex());
#endif
tp->markSmoothWalkDone();
if(verbosity>1)
CkPrintf("[%d] TreePiece %d finished smooth with bucket %d\n",CkMyPe(),
tp->thisIndex,iBucket);
if(verbosity > 1)
ckerr << "TreePiece " << tp->thisIndex << ": My particles are done"
<< endl;
}
}
}
void MarkSmoothCompute::walkDone(State *state) {
// Nothing to do
}
/*
* Functions from PKDGRAV
*/
void Density(GravityParticle *p,int nSmooth, pqSmoothNode *nnList)
{
double ih2,r2,rs,fDensity;
int i;
ih2 = invH2(p);
fDensity = 0.0;
for (i=0;i<nSmooth;++i) {
double fDist2 = nnList[i].fKey;
r2 = fDist2*ih2;
rs = KERNEL(r2);
fDensity += rs*nnList[i].p->mass;
}
p->fDensity = M_1_PI*sqrt(ih2)*ih2*fDensity;
}
int DensitySmoothParams::isSmoothActive(GravityParticle *p)
{
return (p->rung >= activeRung && TYPETest(p, iType));
}
void DensitySmoothParams::initSmoothParticle(GravityParticle *p)
{
p->fDensity = 0.0;
}
void DensitySmoothParams::initSmoothCache(GravityParticle *p)
{
p->fDensity = 0.0;
}
void DensitySmoothParams::combSmoothCache(GravityParticle *p1,
ExternalSmoothParticle *p2) {
p1->fDensity += p2->fDensity;
}
void DensitySmoothParams::fcnSmooth(GravityParticle *p,int nSmooth,
pqSmoothNode *nnList)
{
GravityParticle *q;
double fNorm,ih2,r2,rs;
int i;
ih2 = invH2(p);
fNorm = 0.5*M_1_PI*sqrt(ih2)*ih2;
for (i=0;i<nSmooth;++i) {
double fDist2 = nnList[i].fKey;
r2 = fDist2*ih2;
rs = KERNEL(r2);
rs *= fNorm;
q = nnList[i].p;
p->fDensity += rs*q->mass;
q->fDensity += rs*p->mass;
}
}
![swh spinner](/static/img/swh-spinner.gif)
Computing file changes ...