void octree::direct_dust(tree_structure &tree)
{
if(tree.n_dust == 0) return;
directGrav.set_arg<cl_mem>(0, tree.dust_acc1.p());
directGrav.set_arg<cl_mem>(1, tree.dust_pos.p());
directGrav.set_arg<cl_mem>(2, tree.bodies_Ppos.p());
directGrav.set_arg<int>(3, &tree.n_dust);
directGrav.set_arg<int>(4, &tree.n);
directGrav.set_arg<float>(5, &(this->eps2));
directGrav.set_arg<float4>(6, NULL, 256);
std::vector<size_t> localWork(2), globalWork(2);
localWork[0] = 256; localWork[1] = 1;
globalWork[0] = 256 * ((tree.n_dust + 255) / 256);
globalWork[1] = 1;
directGrav.setWork(globalWork, localWork);
directGrav.execute(gravStream->s()); //First half
}
void octree::gpuSort_32b(my_dev::context &devContext,
my_dev::dev_mem<uint> &srcKeys, my_dev::dev_mem<uint> &srcValues,
my_dev::dev_mem<int> &keysOutput, my_dev::dev_mem<uint> &keysAPing,
my_dev::dev_mem<uint> &valuesOutput,my_dev::dev_mem<uint> &valuesAPing,
int N, int numberOfBits)
{
int bitIdx = 0;
//Step 1, do the count
//Memory that should be alloced outside the function:
setupParams sParam;
sParam.jobs = (N / 64) / 480 ; //64=32*2 2 items per look, 480 is 120*4, number of procs
sParam.blocksWithExtraJobs = (N / 64) % 480;
sParam.extraElements = N % 64;
sParam.extraOffset = N - sParam.extraElements;
sortCount.set_arg<cl_mem>(0, srcKeys.p());
sortCount.set_arg<cl_mem>(1, this->devMemCounts.p());
sortCount.set_arg<uint>(2, &N);
sortCount.set_arg<int>(3, NULL, 128);//smem size
sortCount.set_arg<setupParams>(4, &sParam);
sortCount.set_arg<int>(5, &bitIdx);
vector<size_t> localWork(2), globalWork(2);
globalWork[0] = 32*120; globalWork[1] = 4;
localWork [0] = 32; localWork[1] = 4;
sortCount.setWork(globalWork, localWork);
///////////////
exScanBlock.set_arg<cl_mem>(0, this->devMemCounts.p());
int blocks = 120*4;
exScanBlock.set_arg<int>(1, &blocks);
exScanBlock.set_arg<cl_mem>(2, this->devMemCountsx.p());
exScanBlock.set_arg<int>(3, NULL, 512); //shared memory allocation
globalWork[0] = 512; globalWork[1] = 1;
localWork [0] = 512; localWork [1] = 1;
exScanBlock.setWork(globalWork, localWork);
//////////////
sortMove.set_arg<cl_mem>(0, srcKeys.p());
sortMove.set_arg<cl_mem>(1, keysOutput.p());
sortMove.set_arg<cl_mem>(2, srcValues.p());
sortMove.set_arg<cl_mem>(3, valuesOutput.p());
sortMove.set_arg<cl_mem>(4, this->devMemCounts.p());
sortMove.set_arg<uint>(5, &N);
sortMove.set_arg<uint>(6, NULL, 192); //Dynamic shared memory 128+64 , prefux sum buffer
sortMove.set_arg<uint>(7, NULL, 64*4); //Dynamic shared memory stage buffer
sortMove.set_arg<uint>(8, NULL, 64*4); //Dynamic shared memory stage_values buffer
sortMove.set_arg<setupParams>(9, &sParam);
sortMove.set_arg<int>(10, &bitIdx);
globalWork[0] = 120*32; globalWork[1] = 4;
localWork [0] = 32; localWork [1] = 4;
sortMove.setWork(globalWork, localWork);
bool pingPong = false;
//Execute bitIdx 0
sortCount.execute(execStream->s());
exScanBlock.execute(execStream->s());
sortMove.execute(execStream->s());
//Swap buffers
sortCount.set_arg<cl_mem>(0, keysOutput.p());
sortMove.set_arg<cl_mem>(0, keysOutput.p());
sortMove.set_arg<cl_mem>(1, keysAPing.p());
sortMove.set_arg<cl_mem>(2, valuesOutput.p());
sortMove.set_arg<cl_mem>(3, valuesAPing.p());
//Remaining bits, ping ponging buffers
for(int i=1; i < numberOfBits; i++)
{
bitIdx = i;
sortCount.set_arg<int>(5, &bitIdx);
sortMove.set_arg<int>(10, &bitIdx);
sortCount.execute(execStream->s());
exScanBlock.execute(execStream->s());
sortMove.execute(execStream->s());
//Switch buffers
if(pingPong)
{
sortCount.set_arg<cl_mem>(0, keysOutput.p());
sortMove.set_arg<cl_mem>(0, keysOutput.p());
sortMove.set_arg<cl_mem>(1, keysAPing.p());
sortMove.set_arg<cl_mem>(2, valuesOutput.p());
sortMove.set_arg<cl_mem>(3, valuesAPing.p());
pingPong = false;
}
else
{
sortCount.set_arg<cl_mem>(0, keysAPing.p());
sortMove.set_arg<cl_mem>(0, keysAPing.p());
sortMove.set_arg<cl_mem>(1, keysOutput.p());
sortMove.set_arg<cl_mem>(2, valuesAPing.p());
sortMove.set_arg<cl_mem>(3, valuesOutput.p());
pingPong = true;
}
}
}
//Splits an array of integers, the values in srcValid indicate if a
//value is valid (1 == valid anything else is UNvalid) returns the
//splitted values in the output array (first all valid
//number and then the invalid ones) and the total
//number of valid items is stored in 'count'
void octree::gpuSplit(my_dev::context &devContext,
my_dev::dev_mem<uint> &srcValues,
my_dev::dev_mem<uint> &output,
int N,
int *validCount) // if validCount NULL leave count on device
{
//In the next step we associate the GPU memory with the Kernel arguments
//my_dev::dev_mem<uint> counts(devContext, 512), countx(devContext, 512);
//Memory that should be alloced outside the function:
//devMemCounts and devMemCountsx
// make sure previous reset has finished.
this->devMemCountsx.waitForCopyEvent();
//Kernel configuration parameters
setupParams sParam;
sParam.jobs = (N / 64) / 480 ; //64=32*2 2 items per look, 480 is 120*4, number of procs
sParam.blocksWithExtraJobs = (N / 64) % 480;
sParam.extraElements = N % 64;
sParam.extraOffset = N - sParam.extraElements;
compactCount.set_arg<cl_mem>(0, srcValues.p());
compactCount.set_arg<cl_mem>(1, this->devMemCounts.p());
compactCount.set_arg<uint>(2, &N);
compactCount.set_arg<int>(3, NULL, 128);
compactCount.set_arg<setupParams>(4, &sParam);
compactCount.set_arg<cl_mem>(5, this->devMemCountsx.p());
vector<size_t> localWork(2), globalWork(2);
globalWork[0] = 32*120; globalWork[1] = 4;
localWork [0] = 32; localWork[1] = 4;
compactCount.setWork(globalWork, localWork);
///////////////
exScanBlock.set_arg<cl_mem>(0, this->devMemCounts.p());
int blocks = 120*4;
exScanBlock.set_arg<int>(1, &blocks);
exScanBlock.set_arg<cl_mem>(2, this->devMemCountsx.p());
exScanBlock.set_arg<int>(3, NULL, 512); //shared memory allocation
globalWork[0] = 512; globalWork[1] = 1;
localWork [0] = 512; localWork [1] = 1;
exScanBlock.setWork(globalWork, localWork);
//////////////
splitMove.set_arg<cl_mem>(0, srcValues.p());
splitMove.set_arg<cl_mem>(1, output.p());
splitMove.set_arg<cl_mem>(2, this->devMemCounts.p());
splitMove.set_arg<uint>(3, &N);
splitMove.set_arg<uint>(4, NULL, 192); //Dynamic shared memory
splitMove.set_arg<setupParams>(5, &sParam);
globalWork[0] = 120*32; globalWork[1] = 4;
localWork [0] = 32; localWork [1] = 4;
splitMove.setWork(globalWork, localWork);
////////////////////
compactCount.execute(execStream->s());
exScanBlock.execute(execStream->s());
splitMove.execute(execStream->s());
if (validCount) {
this->devMemCountsx.d2h();
*validCount = this->devMemCountsx[0];
}
}
void octree::compute_properties (tree_structure &tree) {
#if 0
fprintf(stderr,"This file is not up to date anymore! %s\n", __FILE__);
exit(0);
//Computes the tree-properties (size, cm, monopole, quadropole, etc)
//start the kernel for the leaf-type nodes
propsLeaf.set_arg<int>(0, &tree.n_leafs);
propsLeaf.set_arg<cl_mem>(1, tree.leafNodeIdx.p());
propsLeaf.set_arg<cl_mem>(2, tree.node_bodies.p());
propsLeaf.set_arg<cl_mem>(3, tree.bodies_Ppos.p());
// propsLeaf.set_arg<cl_mem>(3, tree.bodies_pos.p());
propsLeaf.set_arg<cl_mem>(4, tree.multipole.p());
propsLeaf.set_arg<cl_mem>(5, tree.nodeLowerBounds.p());
propsLeaf.set_arg<cl_mem>(6, tree.nodeUpperBounds.p());
propsLeaf.set_arg<cl_mem>(7, tree.lowerBounds.p());
propsLeaf.set_arg<cl_mem>(8, tree.upperBounds.p());
propsLeaf.set_arg<cl_mem>(9, tree.bodies_Pvel.p()); //Velocity to get max eps
propsLeaf.setWork(tree.n_leafs, 128);
printf("PropsLeaf: "); propsLeaf.printWorkSize();
propsLeaf.execute();
int temp = tree.n_nodes-tree.n_leafs;
propsNonLeaf.set_arg<int>(0, &temp);
propsNonLeaf.set_arg<cl_mem>(1, tree.leafNodeIdx.p());
propsNonLeaf.set_arg<cl_mem>(2, tree.node_level_list.p());
propsNonLeaf.set_arg<cl_mem>(3, tree.n_children.p());
propsNonLeaf.set_arg<cl_mem>(4, tree.multipole.p());
propsNonLeaf.set_arg<cl_mem>(5, tree.nodeLowerBounds.p());
propsNonLeaf.set_arg<cl_mem>(6, tree.nodeUpperBounds.p());
for(int i=tree.n_levels; i >= 1; i--)
{
propsNonLeaf.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];
propsNonLeaf.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]);
propsNonLeaf.printWorkSize();
}
propsNonLeaf.set_arg<int>(0, &i); //set the level
propsNonLeaf.execute();
}
float theta2 = theta;
propsScaling.set_arg<int>(0, &tree.n_nodes);
propsScaling.set_arg<real4>(1, &tree.corner);
propsScaling.set_arg<cl_mem>(2, tree.multipole.p());
propsScaling.set_arg<cl_mem>(3, tree.nodeLowerBounds.p());
propsScaling.set_arg<cl_mem>(4, tree.nodeUpperBounds.p());
propsScaling.set_arg<cl_mem>(5, tree.n_children.p());
propsScaling.set_arg<cl_mem>(6, tree.node_data.p());
propsScaling.set_arg<float >(7, &theta2);
propsScaling.set_arg<cl_mem>(8, tree.boxSizeInfo.p());
propsScaling.set_arg<cl_mem>(9, tree.boxCenterInfo.p());
propsScaling.setWork(tree.n_nodes, 128);
printf("propsScaling: \t "); propsScaling.printWorkSize();
propsScaling.execute();
//tree.multipole.d2h();
//printf("COM: %f %f %f %f \n",tree.multipole[0].x, tree.multipole[0].y, tree.multipole[0].z, tree.multipole[0].w);
#ifdef USE_CUDA
cuCtxSynchronize();
#else
clFinish(devContext.get_command_queue());
#endif
tree.nodeLowerBounds.d2h();
tree.nodeUpperBounds.d2h();
copyNodeDataToGroupData.set_arg<int>(0, &tree.n_groups);
copyNodeDataToGroupData.set_arg<int>(1, &tree.n_nodes);
copyNodeDataToGroupData.set_arg<cl_mem>(2, tree.node_data.p());
copyNodeDataToGroupData.set_arg<cl_mem>(3, tree.group_data.p());
copyNodeDataToGroupData.set_arg<cl_mem>(4, tree.node_bodies.p());
copyNodeDataToGroupData.set_arg<cl_mem>(5, tree.group_list.p());
copyNodeDataToGroupData.set_arg<cl_mem>(6, tree.boxCenterInfo.p());
copyNodeDataToGroupData.set_arg<cl_mem>(7, tree.boxSizeInfo.p());
copyNodeDataToGroupData.set_arg<cl_mem>(8, tree.groupCenterInfo.p());
copyNodeDataToGroupData.set_arg<cl_mem>(9, tree.groupSizeInfo.p());
copyNodeDataToGroupData.setWork(tree.n_nodes, 128);
printf("copyNodeDataToGroupData: \t "); copyNodeDataToGroupData.printWorkSize();
copyNodeDataToGroupData.execute();
// tree.multipole.d2h();
// testRes.d2h();
// for(int i=0; i < tree.n_nodes; i++)
// for(int i=tree.n_nodes-10; i < tree.n_nodes; i++)
/* for(int i=0; i < 10; i++)
{
fprintf(stderr,"%d\t%f\t%f\t%f\t%f\n", i, tree.multipole[i*3+0].x,tree.multipole[i*3+0].y,tree.multipole[i*3+0].z, tree.multipole[i*3+0].w);
// fprintf(stderr,"%d\t%f\t%f\t%f\t%f\t%f\n", i, tree.multipole[i*3+1].x,tree.multipole[i*3+1].y,tree.multipole[i*3+1].z, tree.multipole[i*3+1].w, testRes[i]);
fprintf(stderr,"%d\t%f\t%f\t%f\t%f\t%f\n", i, tree.multipole[i*3+1].x,tree.multipole[i*3+1].y,tree.multipole[i*3+1].z, tree.multipole[i*3+1].w, 0);
fprintf(stderr,"%d\t%f\t%f\t%f\t%f\n", i, tree.multipole[i*3+2].x,tree.multipole[i*3+2].y,tree.multipole[i*3+2].z, tree.multipole[i*3+2].w);
}
exit(0);
*/
#else
compute_properties_double(tree);
#endif
}
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
}
// If srcValues and buffer are different, then the original values
// are preserved, if they are the same srcValues will be overwritten
void octree::gpuSort(my_dev::context &devContext,
my_dev::dev_mem<uint4> &srcValues,
my_dev::dev_mem<uint4> &output,
my_dev::dev_mem<uint4> &buffer,
int N, int numberOfBits, int subItems,
tree_structure &tree) {
//Extra buffer values
// my_dev::dev_mem<uint> simpleKeys(devContext, N); //Int keys,
// my_dev::dev_mem<uint> permutation(devContext, N); //Permutation values, for sorting the int4 data
// my_dev::dev_mem<int> output32b(devContext, N); //Permutation values, for sorting the int4 data
// my_dev::dev_mem<uint> valuesOutput(devContext, N); //Buffers for the values which are the indexes
my_dev::dev_mem<uint> simpleKeys(devContext); //Int keys,
my_dev::dev_mem<uint> permutation(devContext); //Permutation values, for sorting the int4 data
my_dev::dev_mem<int> output32b(devContext); //Permutation values, for sorting the int4 data
my_dev::dev_mem<uint> valuesOutput(devContext); //Buffers for the values which are the indexes
int prevOffsetSum = getAllignmentOffset(4*N); //The offset of output
simpleKeys.cmalloc_copy(tree.generalBuffer1.get_pinned(),
tree.generalBuffer1.get_flags(),
tree.generalBuffer1.get_devMem(),
&tree.generalBuffer1[8*N], 8*N,
N, prevOffsetSum + getAllignmentOffset(8*N + prevOffsetSum)); //Ofset 8 since we have 2 uint4 before
prevOffsetSum += getAllignmentOffset(8*N + prevOffsetSum);
permutation.cmalloc_copy(tree.generalBuffer1.get_pinned(),
tree.generalBuffer1.get_flags(),
tree.generalBuffer1.get_devMem(),
&tree.generalBuffer1[9*N], 9*N,
N, prevOffsetSum + getAllignmentOffset(9*N + prevOffsetSum)); //N elements after simpleKeys
prevOffsetSum += getAllignmentOffset(9*N + prevOffsetSum);
output32b.cmalloc_copy(tree.generalBuffer1.get_pinned(),
tree.generalBuffer1.get_flags(),
tree.generalBuffer1.get_devMem(),
&tree.generalBuffer1[10*N], 10*N,
N, prevOffsetSum + getAllignmentOffset(10*N + prevOffsetSum)); //N elements after permutation
prevOffsetSum += getAllignmentOffset(10*N + prevOffsetSum);
valuesOutput.cmalloc_copy(tree.generalBuffer1.get_pinned(),
tree.generalBuffer1.get_flags(),
tree.generalBuffer1.get_devMem(),
&tree.generalBuffer1[11*N], 11*N,
N, prevOffsetSum + getAllignmentOffset(11*N + prevOffsetSum)); //N elements after output32b
//Dimensions for the kernels that shuffle and extract data
const int blockSize = 256;
int ng = (N)/blockSize + 1;
int nx = (int)sqrt(ng);
int ny = (ng-1)/nx + 1;
vector<size_t> localWork(2), globalWork(2);
globalWork[0] = nx*blockSize; globalWork[1] = ny;
localWork [0] = blockSize; localWork[1] = 1;
extractInt.setWork(globalWork, localWork);
fillSequence.setWork(globalWork, localWork);
reOrderKeysValues.setWork(globalWork, localWork);
//Idx depends on subitems, z goes first, x last if subitems = 3
//subitems = 3, than idx=2
//subitems = 2, than idx=1
//subitems = 1, than idx=0
//intIdx = subItems-1
int intIdx = subItems-1;
extractInt.set_arg<cl_mem>(0, srcValues.p());
extractInt.set_arg<cl_mem>(1, simpleKeys.p());
extractInt.set_arg<uint>(2, &N);
extractInt.set_arg<int>(3, &intIdx);//bit idx
fillSequence.set_arg<cl_mem>(0, permutation.p());
fillSequence.set_arg<uint>(1, &N);
reOrderKeysValues.set_arg<cl_mem>(0, srcValues.p());
reOrderKeysValues.set_arg<cl_mem>(1, output.p());
reOrderKeysValues.set_arg<cl_mem>(2, valuesOutput.p());
reOrderKeysValues.set_arg<uint>(3, &N);
extractInt.execute();
fillSequence.execute();
//Now sort the first 32bit keys
//Using 32bit sort with key and value seperated
gpuSort_32b(devContext,
simpleKeys, permutation,
// output32b, aPing32b,
output32b, simpleKeys,
// valuesOutput,valuesAPing,
valuesOutput,permutation,
// count,
N, 32);
//Now reorder the main keys
//Use output as the new output/src value thing buffer
reOrderKeysValues.execute();
if(subItems == 1)
{
//Only doing one 32bit sort. Data is already in output so done
return;
}
//2nd set of 32bit keys
//Idx depends on subitems, z goes first, x last if subitems = 3
//subitems = 3, than idx=1
//subitems = 2, than idx=0
//subitems = 1, completed previous round
//intIdx = subItems-2
intIdx = subItems-2;
extractInt.set_arg<cl_mem>(0, output.p());
extractInt.set_arg<int>(3, &intIdx);//smem size
reOrderKeysValues.set_arg<cl_mem>(0, output.p());
reOrderKeysValues.set_arg<cl_mem>(1, buffer.p());
extractInt.execute();
fillSequence.execute();
//Now sort the 2nd 32bit keys
//Using 32bit sort with key and value seperated
gpuSort_32b(devContext,
simpleKeys, permutation,
output32b, simpleKeys,
// output32b, aPing32b,
// valuesOutput,valuesAPing,
valuesOutput,permutation,
//count,
N, 32);
reOrderKeysValues.execute();
if(subItems == 2)
{
//Doing two 32bit sorts. Data is in buffer
//so move the data from buffer to output
output.copy(buffer, buffer.get_size());
return;
}
//3th set of 32bit keys
//Idx depends on subitems, z goes first, x last if subitems = 3
//subitems = 3, than idx=0
//subitems = 2, completed previous round
//subitems = 1, completed previous round
//intIdx = subItems-2
intIdx = 0;
extractInt.set_arg<cl_mem>(0, buffer.p());
extractInt.set_arg<int>(3, &intIdx);//integer idx
reOrderKeysValues.set_arg<cl_mem>(0, buffer.p());
reOrderKeysValues.set_arg<cl_mem>(1, output.p());
extractInt.execute();
fillSequence.execute();
//Now sort the 32bit keys
//Using int2 with key and value combined
//See sortArray4
//Using key and value in a seperate array
//Now sort the 2nd 32bit keys
//Using 32bit sort with key and value seperated
gpuSort_32b(devContext,
simpleKeys, permutation,
output32b, simpleKeys,
// output32b, aPing32b,
// valuesOutput,valuesAPing,
valuesOutput,permutation,
//count,
N, 32);
reOrderKeysValues.execute();
clFinish(devContext.get_command_queue());
// fprintf(stderr, "sortArray2 done in %g sec (Without memory alloc & compilation) \n", get_time() - t0);
}
//Splits an array of integers, the values in srcValid indicate if a
//value is valid (1 == valid anything else is UNvalid) returns the
//splitted values in the output array (first all valid
//number and then the invalid ones) and the total
//number of valid items is stored in 'count'
void octree::gpuSplit(my_dev::context &devContext,
my_dev::dev_mem<uint> &srcValues,
my_dev::dev_mem<uint> &output,
int N, int *validCount)
{
// In the next step we associate the GPU memory with the Kernel arguments
// my_dev::dev_mem<uint> counts(devContext, 512), countx(devContext, 512);
//Memory that should be alloced outside the function:
//devMemCounts and devMemCountsx
//Kernel configuration parameters
setupParams sParam;
sParam.jobs = (N / 64) / 480 ; //64=32*2 2 items per look, 480 is 120*4, number of procs
sParam.blocksWithExtraJobs = (N / 64) % 480;
sParam.extraElements = N % 64;
sParam.extraOffset = N - sParam.extraElements;
// printf("Param info: %d %d %d %d \n", sParam.jobs, sParam.blocksWithExtraJobs, sParam.extraElements, sParam.extraOffset);
compactCount.set_arg<cl_mem>(0, srcValues.p());
compactCount.set_arg<cl_mem>(1, this->devMemCounts.p());
compactCount.set_arg<uint>(2, &N);
compactCount.set_arg<int>(3, NULL, 128);
compactCount.set_arg<setupParams>(4, &sParam);
vector<size_t> localWork(2), globalWork(2);
globalWork[0] = 32*120; globalWork[1] = 4;
localWork [0] = 32; localWork[1] = 4;
compactCount.setWork(globalWork, localWork);
///////////////
exScanBlock.set_arg<cl_mem>(0, this->devMemCounts.p());
int blocks = 120*4;
exScanBlock.set_arg<int>(1, &blocks);
exScanBlock.set_arg<cl_mem>(2, this->devMemCountsx.p());
exScanBlock.set_arg<int>(3, NULL, 512); //shared memory allocation
globalWork[0] = 512; globalWork[1] = 1;
localWork [0] = 512; localWork [1] = 1;
exScanBlock.setWork(globalWork, localWork);
//////////////
splitMove.set_arg<cl_mem>(0, srcValues.p());
splitMove.set_arg<cl_mem>(1, output.p());
splitMove.set_arg<cl_mem>(2, this->devMemCounts.p());
splitMove.set_arg<uint>(3, &N);
splitMove.set_arg<uint>(4, NULL, 192); //Dynamic shared memory
splitMove.set_arg<setupParams>(5, &sParam);
globalWork[0] = 120*32; globalWork[1] = 4;
localWork [0] = 32; localWork [1] = 4;
splitMove.setWork(globalWork, localWork);
////////////////////
compactCount.execute();
// exit(0);
// counts.d2h();
// for(int i=0; i < 482; i++)
// {
// printf("%d\t%d\n", i, counts[i]);
// }
//
exScanBlock.execute();
splitMove.execute();
//TODO fix the damn clFinish function
#ifdef USE_CUDA
cuCtxSynchronize();
#else
clFinish(devContext.get_command_queue());
#endif
this->devMemCountsx.d2h();
*validCount = this->devMemCountsx[0];
//printf("Total number of valid items: %d \n", countx[0]);
}
void octree::compute_properties(tree_structure &tree, my_dev::dev_mem<float4> &bodies_pos, int n_bodies) {
/*****************************************************
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 -> 2*3*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
(Note that generalBuffer might be larger because of tree-walk stack)
*****************************************************/
if(10*tree.n_nodes > 3*tree.n)
{
#ifdef _DEBUG_PRINT_
fprintf(stderr, "Resizeing the generalBuffer1 \n");
#endif
tree.generalBuffer1.cresize(10*tree.n_nodes*4, false);
}
my_dev::dev_mem<double4> multipoleD(devContext);
my_dev::dev_mem<real4> nodeLowerBounds(devContext); //Lower bounds used for scaling? TODO
my_dev::dev_mem<real4> nodeUpperBounds(devContext); //Upper bounds used for scaling? TODO
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
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
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, bodies_pos.p());
propsLeafD.set_arg<cl_mem>(4, multipoleD.p());
propsLeafD.set_arg<cl_mem>(5, nodeLowerBounds.p());
propsLeafD.set_arg<cl_mem>(6, nodeUpperBounds.p());
// propsLeafD.set_arg<cl_mem>(7, tree.bodies_Pvel.p()); //Velocity to get max eps
propsLeafD.setWork(tree.n_leafs, 128);
#ifdef _DEBUG_PRINT_
printf("PropsLeaf: ");
propsLeafD.printWorkSize();
#endif
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, nodeLowerBounds.p());
propsNonLeafD.set_arg<cl_mem>(6, 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);
#ifdef _DEBUG_PRINT_
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]);
#endif
propsNonLeafD.printWorkSize();
}
propsNonLeafD.set_arg<int>(0, &i); //set the level
propsNonLeafD.execute();
}
propsScalingD.set_arg<int>(0, &tree.n_nodes);
propsScalingD.set_arg<cl_mem>(1, multipoleD.p());
propsScalingD.set_arg<cl_mem>(2, nodeLowerBounds.p());
propsScalingD.set_arg<cl_mem>(3, nodeUpperBounds.p());
propsScalingD.set_arg<cl_mem>(4, tree.n_children.p());
propsScalingD.set_arg<cl_mem>(5, tree.multipole.p());
propsScalingD.set_arg<float >(6, &theta);
propsScalingD.set_arg<cl_mem>(7, tree.boxSizeInfo.p());
propsScalingD.set_arg<cl_mem>(8, tree.boxCenterInfo.p());
propsScalingD.set_arg<cl_mem>(9, tree.node_bodies.p());
propsScalingD.setWork(tree.n_nodes, 128);
#ifdef _DEBUG_PRINT_
printf("propsScaling: \t ");
propsScalingD.printWorkSize();
#endif
propsScalingD.execute();
#if 0
#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);
#endif
}
