void octree::compute_properties_double(tree_structure &tree) {
/*****************************************************
Assign the memory buffers, note that we check the size first
and if needed we increase the size of the generalBuffer1
Size required:
- multipoleD -> double4*3_n_nodes -> 6*n_nodes*uint4
- lower/upperbounds -> 2*n_nodes*uint4
- node lower/upper -> 2*n_nodes*uint4
- SUM: 10*n_nodes*uint4
- generalBuffer1 has default size: 3*N*uint4
check if 10*n_nodes < 3*N if so realloc
*****************************************************/
if(10*tree.n_nodes > 3*tree.n)
{
fprintf(stderr, "Resizeing the generalBuffer1 \n");
tree.generalBuffer1.cresize(8*tree.n_nodes*4, false);
}
my_dev::dev_mem<double4> multipoleD(devContext);
multipoleD.cmalloc_copy(tree.generalBuffer1.get_pinned(),
tree.generalBuffer1.get_flags(),
tree.generalBuffer1.get_devMem(),
&tree.generalBuffer1[0], 0,
3*tree.n_nodes, getAllignmentOffset(0));
//Offset is in uint, so: double4 = 8uint*3*n_nodes
tree.nodeLowerBounds.cmalloc_copy(tree.generalBuffer1.get_pinned(),
tree.generalBuffer1.get_flags(),
tree.generalBuffer1.get_devMem(),
&tree.generalBuffer1[8*3*tree.n_nodes], 8*3*tree.n_nodes,
tree.n_nodes, getAllignmentOffset(8*3*tree.n_nodes));
int prevOffsetSum = getAllignmentOffset(8*3*tree.n_nodes); //The offset of output
tree.nodeUpperBounds.cmalloc_copy(tree.generalBuffer1.get_pinned(),
tree.generalBuffer1.get_flags(),
tree.generalBuffer1.get_devMem(),
&tree.generalBuffer1[8*3*tree.n_nodes + 4*tree.n_nodes],
8*3*tree.n_nodes + 4*tree.n_nodes,
tree.n_nodes,
prevOffsetSum + getAllignmentOffset(8*3*tree.n_nodes + 4*tree.n_nodes + prevOffsetSum));
//Computes the tree-properties (size, cm, monopole, quadropole, etc)
//start the kernel for the leaf-type nodes
propsLeafD.set_arg<int>(0, &tree.n_leafs);
propsLeafD.set_arg<cl_mem>(1, tree.leafNodeIdx.p());
propsLeafD.set_arg<cl_mem>(2, tree.node_bodies.p());
propsLeafD.set_arg<cl_mem>(3, tree.bodies_Ppos.p());
propsLeafD.set_arg<cl_mem>(4, multipoleD.p());
propsLeafD.set_arg<cl_mem>(5, tree.nodeLowerBounds.p());
propsLeafD.set_arg<cl_mem>(6, tree.nodeUpperBounds.p());
propsLeafD.set_arg<cl_mem>(7, tree.bodies_Pvel.p()); //Velocity to get max eps
propsLeafD.setWork(tree.n_leafs, 128);
printf("PropsLeaf: "); propsLeafD.printWorkSize();
propsLeafD.execute();
int temp = tree.n_nodes-tree.n_leafs;
propsNonLeafD.set_arg<int>(0, &temp);
propsNonLeafD.set_arg<cl_mem>(1, tree.leafNodeIdx.p());
propsNonLeafD.set_arg<cl_mem>(2, tree.node_level_list.p());
propsNonLeafD.set_arg<cl_mem>(3, tree.n_children.p());
propsNonLeafD.set_arg<cl_mem>(4, multipoleD.p());
propsNonLeafD.set_arg<cl_mem>(5, tree.nodeLowerBounds.p());
propsNonLeafD.set_arg<cl_mem>(6, tree.nodeUpperBounds.p());
//Work from the bottom up
for(int i=tree.n_levels; i >= 1; i--)
{
propsNonLeafD.set_arg<int>(0, &i);
{
vector<size_t> localWork(2), globalWork(2);
int totalOnThisLevel;
totalOnThisLevel = tree.node_level_list[i]-tree.node_level_list[i-1];
propsNonLeafD.setWork(totalOnThisLevel, 128);
printf("PropsNonLeaf, nodes on level %d : %d (start: %d end: %d) , config: \t", i, totalOnThisLevel,
tree.node_level_list[i-1], tree.node_level_list[i]);
propsNonLeafD.printWorkSize();
}
propsNonLeafD.set_arg<int>(0, &i); //set the level
propsNonLeafD.execute();
}
float theta2 = theta;
propsScalingD.set_arg<int>(0, &tree.n_nodes);
propsScalingD.set_arg<real4>(1, &tree.corner);
propsScalingD.set_arg<cl_mem>(2, multipoleD.p());
propsScalingD.set_arg<cl_mem>(3, tree.nodeLowerBounds.p());
propsScalingD.set_arg<cl_mem>(4, tree.nodeUpperBounds.p());
propsScalingD.set_arg<cl_mem>(5, tree.n_children.p());
propsScalingD.set_arg<cl_mem>(6, tree.multipole.p());
propsScalingD.set_arg<float >(7, &theta2);
propsScalingD.set_arg<cl_mem>(8, tree.boxSizeInfo.p());
propsScalingD.set_arg<cl_mem>(9, tree.boxCenterInfo.p());
propsScalingD.set_arg<cl_mem>(10, tree.node_bodies.p());
propsScalingD.setWork(tree.n_nodes, 128);
printf("propsScaling: \t "); propsScalingD.printWorkSize();
propsScalingD.execute();
#ifdef INDSOFT
//If we use individual softening we need to get the max softening value
//to be broadcasted during the exchange of the LET boundaries.
//Only copy the root node that contains the max value
my_dev::dev_stream memCpyStream;
tree.multipole.d2h(3, false, memCpyStream.s());
#endif
//Set the group properties, note that it is not based on the nodes anymore
//but on self created groups based on particle order setPHGroupData
copyNodeDataToGroupData.set_arg<int>(0, &tree.n_groups);
copyNodeDataToGroupData.set_arg<int>(1, &tree.n);
copyNodeDataToGroupData.set_arg<cl_mem>(2, tree.bodies_Ppos.p());
copyNodeDataToGroupData.set_arg<cl_mem>(3, tree.group_list_test.p());
copyNodeDataToGroupData.set_arg<cl_mem>(4, tree.groupCenterInfo.p());
copyNodeDataToGroupData.set_arg<cl_mem>(5, tree.groupSizeInfo.p());
copyNodeDataToGroupData.setWork(-1, NCRIT, tree.n_groups);
copyNodeDataToGroupData.printWorkSize();
copyNodeDataToGroupData.execute();
#ifdef INDSOFT
memCpyStream.sync();
this->maxLocalEps = tree.multipole[0*3 + 1].w; //Softening value
#else
#endif
//Get the local domain boundary based on group positions and sizes
real4 r_min, r_max;
getBoundariesGroups(tree, r_min, r_max);
#if 0
//Write the tree structure to file
string nodeFileName = "fullTreeStructure.txt";
char fileName[256];
sprintf(fileName, "fullTreeStructure-%d.txt", mpiGetRank());
ofstream nodeFile;
//nodeFile.open(nodeFileName.c_str());
nodeFile.open(fileName);
tree.multipole.d2h();
tree.boxSizeInfo.d2h();
tree.boxCenterInfo.d2h();
for(int i=0; i < tree.n_nodes; i++)
{
//nodeFile << i << "\t" << tree.boxCenterInfo[i].x << "\t" << tree.boxCenterInfo[i].y;
//nodeFile << "\t" << 2*tree.boxSizeInfo[i].x << "\t" << 2*tree.boxSizeInfo[i].y << "\t";
nodeFile << i << "\t" << tree.boxCenterInfo[i].x-tree.boxSizeInfo[i].x << "\t" << tree.boxCenterInfo[i].y-tree.boxSizeInfo[i].y;
nodeFile << "\t" << tree.boxCenterInfo[i].x+tree.boxSizeInfo[i].x << "\t" << tree.boxCenterInfo[i].y+tree.boxSizeInfo[i].y << "\t";
nodeFile << tree.multipole[i*3+0].x << "\t" << tree.multipole[i*3+0].w << "\n";
}
nodeFile.close();
sprintf(fileName, "fullTreeStructureParticles-%d.txt", mpiGetRank());
ofstream partFile;
partFile.open(fileName);
tree.bodies_pos.d2h();
for(int i=0; i < tree.n; i++)
{
float4 pos = tree.bodies_pos[i];
partFile << i << "\t" << pos.x << "\t" << pos.y << "\t" << pos.z << endl;
}
partFile.close();
#endif
}
void octree::approximate_dust(tree_structure &tree)
{
if(tree.n_dust == 0) return;
uint2 node_begend;
int level_start = 2;
node_begend.x = tree.level_list[level_start].x;
node_begend.y = tree.level_list[level_start].y;
//Reset the active particles
tree.active_dust_list.zeroMem();
//Set the kernel parameters, many!
approxGrav.set_arg<int>(0, &tree.n_dust_groups);
approxGrav.set_arg<int>(1, &tree.n_dust);
approxGrav.set_arg<float>(2, &(this->eps2));
approxGrav.set_arg<uint2>(3, &node_begend);
approxGrav.set_arg<cl_mem>(4, tree.activeDustGrouplist.p());
approxGrav.set_arg<cl_mem>(5, tree.bodies_Ppos.p()); //Bodies from the tree
approxGrav.set_arg<cl_mem>(6, tree.multipole.p());
approxGrav.set_arg<cl_mem>(7, tree.dust_acc1.p());
approxGrav.set_arg<cl_mem>(8, tree.dust_pos.p()); //Dust bodies
approxGrav.set_arg<cl_mem>(9, tree.dust_ngb.p());
approxGrav.set_arg<cl_mem>(10, tree.active_dust_list.p());
approxGrav.set_arg<cl_mem>(11, tree.dust_interactions.p());
approxGrav.set_arg<cl_mem>(12, tree.boxSizeInfo.p());
approxGrav.set_arg<cl_mem>(13, tree.dust_groupSizeInfo.p());
approxGrav.set_arg<cl_mem>(14, tree.boxCenterInfo.p());
approxGrav.set_arg<cl_mem>(15, tree.dust_groupCenterInfo.p());
approxGrav.set_arg<cl_mem>(16, tree.dust_vel.p());
approxGrav.set_arg<cl_mem>(17, tree.generalBuffer1.p()); //Instead of using Local memory
approxGrav.set_arg<real4>(18, tree.boxSizeInfo, 4, "texNodeSize");
approxGrav.set_arg<real4>(19, tree.boxCenterInfo, 4, "texNodeCenter");
approxGrav.set_arg<real4>(20, tree.multipole, 4, "texMultipole");
approxGrav.set_arg<real4>(21, tree.dust_pos, 4, "texBody");
approxGrav.setWork(-1, NTHREAD, nBlocksForTreeWalk);
approxGrav.execute(gravStream->s()); //First half
//Print interaction statistics
#if 0
tree.dust_interactions.d2h();
long long directSum = 0;
long long apprSum = 0;
long long directSum2 = 0;
long long apprSum2 = 0;
int maxDir = -1;
int maxAppr = -1;
for(int i=0; i < tree.n_dust; i++)
{
apprSum += tree.dust_interactions[i].x;
directSum += tree.dust_interactions[i].y;
maxAppr = max(maxAppr,tree.dust_interactions[i].x);
maxDir = max(maxDir,tree.dust_interactions[i].y);
apprSum2 += tree.dust_interactions[i].x*tree.dust_interactions[i].x;
directSum2 += tree.dust_interactions[i].y*tree.dust_interactions[i].y;
}
cout << "DUST Interaction at (rank= " << mpiGetRank() << " ) iter: " << iter << "\tdirect: " << directSum << "\tappr: " << apprSum << "\t";
cout << "avg dir: " << directSum / tree.n_dust << "\tavg appr: " << apprSum / tree.n_dust << "\tMaxdir: " << maxDir << "\tmaxAppr: " << maxAppr << endl;
cout << "sigma dir: " << sqrt((directSum2 - directSum)/ tree.n_dust) << "\tsigma appr: " << std::sqrt((apprSum2 - apprSum) / tree.n_dust) << endl;
#endif
/*
tree.dust_acc1.d2h();
tree.dust_pos.d2h();
for(int i=0; i < tree.n_dust; i++)
{
fprintf(stderr,"%d\t%f %f %f %f \t %f %f %f\n", i,
tree.dust_acc1[i].x,
tree.dust_acc1[i].y,
tree.dust_acc1[i].z,
tree.dust_acc1[i].w,
tree.dust_pos[i].x,
tree.dust_pos[i].y,
tree.dust_pos[i].z);
}
*/
//exit(0);
}
void octree::build (tree_structure &tree) {
int level = 0;
int validCount = 0;
int offset = 0;
/******** load kernels **********/
/******** create memory buffers **********/
my_dev::dev_mem<uint> validList(devContext);
my_dev::dev_mem<uint> compactList(devContext);
validList.cmalloc_copy(tree.generalBuffer1.get_pinned(),
tree.generalBuffer1.get_flags(),
tree.generalBuffer1.get_devMem(),
&tree.generalBuffer1[0], 0,
tree.n*2, getAllignmentOffset(0));
validList.zeroMem();
compactList.cmalloc_copy(tree.generalBuffer1.get_pinned(),
tree.generalBuffer1.get_flags(),
tree.generalBuffer1.get_devMem(),
&tree.generalBuffer1[tree.n*2], tree.n*2,
tree.n*2, getAllignmentOffset(tree.n*2));
/******** set kernels parameters **********/
build_key_list.set_arg<cl_mem>(0, tree.bodies_key.p());
build_key_list.set_arg<cl_mem>(1, tree.bodies_Ppos.p());
build_key_list.set_arg<int>(2, &tree.n);
build_key_list.set_arg<real4>(3, &tree.corner);
build_key_list.setWork(tree.n, 128);
build_valid_list.set_arg<int>(0, &tree.n);
build_valid_list.set_arg<int>(1, &level);
build_valid_list.set_arg<cl_mem>(2, tree.bodies_key.p());
build_valid_list.set_arg<cl_mem>(3, validList.p());
build_valid_list.setWork(tree.n, 128);
build_nodes.set_arg<int>(0, &level);
build_nodes.set_arg<int>(1, &validCount);
build_nodes.set_arg<int>(2, &offset);
build_nodes.set_arg<cl_mem>(3, compactList.p());
build_nodes.set_arg<cl_mem>(4, tree.bodies_key.p());
build_nodes.set_arg<cl_mem>(5, tree.node_key.p());
build_nodes.set_arg<cl_mem>(6, tree.n_children.p());
build_nodes.set_arg<cl_mem>(7, tree.node_bodies.p());
link_tree.set_arg<int>(0, &offset);
link_tree.set_arg<cl_mem>(1, tree.n_children.p());
link_tree.set_arg<cl_mem>(2, tree.node_bodies.p());
link_tree.set_arg<cl_mem>(3, tree.bodies_Ppos.p());
link_tree.set_arg<real4>(4, &tree.corner);
link_tree.set_arg<cl_mem>(5, tree.level_list.p());
link_tree.set_arg<cl_mem>(6, validList.p());
link_tree.set_arg<cl_mem>(7, tree.node_key.p());
link_tree.set_arg<cl_mem>(8, tree.bodies_key.p());
link_tree.set_arg<int>(9, &level);
/********** build list of keys ********/
build_key_list.execute();
/****** build the levels *********/
int nodeSum = 0;
for (level = 0; level < MAXLEVELS; level++) {
// mark bodies to be combined into nodes
build_valid_list.set_arg<int>(1, &level);
build_valid_list.execute();
//gpuCompact to get number of created nodes
gpuCompact(devContext, validList, compactList, tree.n*2, &validCount);
nodeSum += validCount / 2;
printf("ValidCount (%d): %d \tSum: %d Offset: %d\n", mpiGetRank(), validCount, nodeSum, offset);
validCount /= 2;
if (validCount == 0) break;
// asssemble nodes
build_nodes.setWork(validCount, 128);
build_nodes.set_arg<int>(0, &level);
build_nodes.set_arg<int>(1, &validCount);
build_nodes.set_arg<int>(2, &offset);
build_nodes.execute();
tree.level_list[level] = (uint2){offset, offset + validCount};
offset += validCount;
} //end for lvl
//Put the last level + 1 index to 0,0
//so we dont need an extra if statement in the linking phase
tree.level_list[level] = (uint2){0, 0};
tree.level_list.h2d();
int n_nodes = offset;
tree.n_nodes = n_nodes;
/***** Link the tree ******/
link_tree.set_arg<int>(0, &offset); //Offset=number of nodes
link_tree.set_arg<int>(9, &level); //level=highest number of levels
//The maximum number of levels that can be used is MAXLEVEl
//if max level is larger than that the program will exit
printf("Max level : %d \n", level);
if(level >= MAXLEVELS)
{
cerr << "The tree has become too deep, the program will exit. \n";
cerr << "Consider the removal of far away particles to prevent a too large box. \n";
maxlevels_exceeded = true;
return;
//exit(0);
}
link_tree.setWork(n_nodes, 128);
printf("Link_tree: "); link_tree.printWorkSize();
tree.n_levels = level-1;
for(int i=0; i < level; i++)
printf("%d\t%d\t%d\n", i, tree.level_list[i].x, tree.level_list[i].y);
//Link the tree
link_tree.execute();
//After executing link_tree, the id_list contains for each node
//the ID of its parent.
//Valid_list contains for each node if its a leaf (valid) or a normal
//node -> non_valid
//Execute a split on the validList to get seperate id lists
//for the leafs and nodes. Used when computing multipoles
tree.leafNodeIdx.cmalloc(tree.n_nodes , false);
//Split the leaf ids and non-leaf node ids
gpuSplit(devContext, validList, tree.leafNodeIdx, tree.n_nodes, &tree.n_leafs);
printf("Total nodes: %d N_leafs: %d non-leafs: %d \n", tree.n_nodes, tree.n_leafs, tree.n_nodes - tree.n_leafs);
build_level_list.set_arg<int>(0, &tree.n_nodes);
build_level_list.set_arg<int>(1, &tree.n_leafs);
build_level_list.set_arg<cl_mem>(2, tree.leafNodeIdx.p());
build_level_list.set_arg<cl_mem>(3, tree.node_bodies.p());
build_level_list.set_arg<cl_mem>(4, validList.p());
build_level_list.setWork(tree.n_nodes-tree.n_leafs, 128);
validList.zeroMem();
//Build the level list based on the leafIdx list
//required for easy access in the compute node properties
build_level_list.execute();
tree.node_level_list.cmalloc(level*2 , false);
int levelThing;
gpuCompact(devContext, validList, tree.node_level_list,
2*(tree.n_nodes-tree.n_leafs), &levelThing);
tree.node_level_list.d2h();
//We only care about end positions, so compress the list:
int j=0;
for(int i=0; i < levelThing; i+=2, j++)
tree.node_level_list[j] = tree.node_level_list[i];
tree.node_level_list[j] =tree.node_level_list[levelThing-1]+1; //Add 1 to make it the end position
levelThing = j+1;
tree.node_level_list.h2d();
printf("Finished level list \n");
for(int i=0; i < levelThing; i++)
{
printf("node_level_list: %d \t%d\n", i, tree.node_level_list[i]);
}
///****** Start building the particle groups *******///////
//Compute the box size, the max length of one of the sides of the rectangle
real size = fmax(fabs(rMaxLocalTree.z - rMinLocalTree.z),
fmax(fabs(rMaxLocalTree.y - rMinLocalTree.y),
fabs(rMaxLocalTree.x - rMinLocalTree.x)));
real dist = ((rMaxLocalTree.z - rMinLocalTree.z) * (rMaxLocalTree.z - rMinLocalTree.z) +
(rMaxLocalTree.y - rMinLocalTree.y) * (rMaxLocalTree.y - rMinLocalTree.y) +
(rMaxLocalTree.x - rMinLocalTree.x) * (rMaxLocalTree.x - rMinLocalTree.x));
float maxDist = sqrt(dist) / 10;
maxDist *= maxDist; //Square since we dont do sqrt on device
fprintf(stderr,"Box max size: %f en max dist: %f \t %f en %f \n", size, dist, sqrt(dist), maxDist);
//maxDist = 50;
validList.zeroMem();
//The newest group creation method!
define_groups.set_arg<int>(0, &tree.n);
define_groups.set_arg<cl_mem>(1, validList.p());
define_groups.set_arg<cl_mem>(2, tree.bodies_Ppos.p());
define_groups.set_arg<float>(3, &maxDist);
define_groups.setWork(tree.n, 128);
define_groups.execute();
//gpuCompact
gpuCompact(devContext, validList, compactList, tree.n*2, &validCount);
printf("Found number of groups: %d \n", validCount/2);
tree.n_groups = validCount/2;
//Now compact validList to get the list of group ids
tree.group_list_test.cmalloc(tree.n_groups , false);
store_groups.set_arg<int>(0, &tree.n);
store_groups.set_arg<int>(1, &tree.n_groups);
store_groups.set_arg<cl_mem>(2, compactList.p());
store_groups.set_arg<cl_mem>(3, tree.body2group_list.p());
store_groups.set_arg<cl_mem>(4, tree.group_list_test.p());
store_groups.setWork(-1, NCRIT, tree.n_groups);
store_groups.execute();
//Memory allocation for the valid group lists
if(tree.active_group_list.get_size() > 0)
{
tree.active_group_list.cresize(tree.n_groups, false);
tree.activeGrpList.cresize(tree.n_groups, false);
}
else
{
tree.active_group_list.cmalloc(tree.n_groups, false);
tree.activeGrpList.cmalloc(tree.n_groups, false);
}
printf("Tree built complete!\n");
/*************************/
}
