void sort_parallel(int a[], int first, int last, int threads) {
if (threads <= 1 || (last - first) <= 100) {
sort_serial(a, first, last);
} else {
int middle = first + (last - first) / 2;
#pragma omp parallel sections
{
#pragma omp section
sort_parallel(a, first, middle, threads / 2 + 1);
#pragma omp section
sort_parallel(a, middle, last, threads / 2 + 1);
}
merge(a, first, middle, last);
}
}
int main(int argc, char **argv) {
FILE *f = fopen(INPUT_FILE, "r");
if (f == NULL) {
printf("Failed to open input file: %s\n", strerror(errno));
goto exit;
}
int size = 0;
while (!feof(f)) {
if (size > MAX_SIZE) {
printf("Too much numbers\n");
break;
}
if (fscanf(f, "%d", &(array[size])) <= 0) {
break;
}
size++;
}
printf("Read %d numbers\n", size);
int num_of_threads = 8; //omp_get_num_threads();
omp_set_num_threads(num_of_threads);
printf("Sorting numbers with %d threads\n", num_of_threads);
sort_parallel(array, 0, size, num_of_threads);
sort_serial(array, 0, size);
save_result(array, size);
fclose(f);
exit:
return 0;
}
void FMEMultipoleKernel::quadtreeConstruction(ArrayPartition& pointPartition)
{
FMELocalContext* localContext = m_pLocalContext;
FMEGlobalContext* globalContext = m_pGlobalContext;
LinearQuadtree& tree = *globalContext->pQuadtree;
// precompute the bounding box for the quadtree points from the graph nodes
for_loop(pointPartition, min_max_x_function(localContext));
for_loop(pointPartition, min_max_y_function(localContext));
// wait until the thread's bounding box is computed
sync();
// let the main thread computed the bounding box of the bounding boxes
if (isMainThread())
{
globalContext->min_x = globalContext->pLocalContext[0]->min_x;
globalContext->min_y = globalContext->pLocalContext[0]->min_y;
globalContext->max_x = globalContext->pLocalContext[0]->max_x;
globalContext->max_y = globalContext->pLocalContext[0]->max_y;
for (__uint32 j=1; j < numThreads(); j++)
{
globalContext->min_x = min(globalContext->min_x, globalContext->pLocalContext[j]->min_x);
globalContext->min_y = min(globalContext->min_y, globalContext->pLocalContext[j]->min_y);
globalContext->max_x = max(globalContext->max_x, globalContext->pLocalContext[j]->max_x);
globalContext->max_y = max(globalContext->max_y, globalContext->pLocalContext[j]->max_y);
};
tree.init(globalContext->min_x, globalContext->min_y, globalContext->max_x, globalContext->max_y);
globalContext->coolDown *= 0.999f;
tree.clear();
};
// wait because the morton number computation needs the bounding box
sync();
// udpate morton number to prepare them for sorting
for_loop(pointPartition, LQMortonFunctor(localContext));
// wait so we can sort them by morton number
sync();
#ifdef OGDF_FME_PARALLEL_QUADTREE_SORT
// use a simple parallel sorting algorithm
LinearQuadtree::LQPoint* points = tree.pointArray();
sort_parallel(points, tree.numberOfPoints(), LQPointComparer);
#else
if (isMainThread())
{
LinearQuadtree::LQPoint* points = tree.pointArray();
sort_single(points, tree.numberOfPoints(), LQPointComparer);
};
#endif
// wait because the quadtree builder needs the sorted order
sync();
// if not a parallel run, we can do the easy way
if (isSingleThreaded())
{
LinearQuadtreeBuilder builder(tree);
// prepare the tree
builder.prepareTree();
// and link it
builder.build();
LQPartitioner partitioner( localContext );
partitioner.partition();
} else // the more difficult part
{
// snap the left point of the interval of the thread to the first in the cell
LinearQuadtree::PointID beginPoint = tree.findFirstPointInCell(pointPartition.begin);
LinearQuadtree::PointID endPoint_plus_one;
// if this thread is the last one, no snapping required for the right point
if (threadNr()==numThreads()-1)
endPoint_plus_one = tree.numberOfPoints();
else // find the left point of the next thread
endPoint_plus_one = tree.findFirstPointInCell(pointPartition.end+1);
// and calculate the number of points to prepare
__uint32 numPointsToPrepare = endPoint_plus_one - beginPoint;
// now we can prepare the snapped interval
LinearQuadtreeBuilder builder(tree);
// this function prepares the tree from begin point to endPoint_plus_one-1 (EXCLUDING endPoint_plus_one)
builder.prepareTree(beginPoint, endPoint_plus_one);
// save the start, end and count of the inner node chain in the context
localContext->firstInnerNode = builder.firstInner;
localContext->lastInnerNode = builder.lastInner;
localContext->numInnerNodes = builder.numInnerNodes;
// save the start, end and count of the leaf node chain in the context
localContext->firstLeaf = builder.firstLeaf;
localContext->lastLeaf = builder.lastLeaf;
localContext->numLeaves = builder.numLeaves;
// wait until all are finished
sync();
// now the main thread has to link the tree
if (isMainThread())
{
// with his own builder
LinearQuadtreeBuilder sbuilder(tree);
// first we need the complete chain data
sbuilder.firstInner = globalContext->pLocalContext[0]->firstInnerNode;
sbuilder.firstLeaf = globalContext->pLocalContext[0]->firstLeaf;
sbuilder.numInnerNodes = globalContext->pLocalContext[0]->numInnerNodes;
sbuilder.numLeaves = globalContext->pLocalContext[0]->numLeaves;
for (__uint32 j=1; j < numThreads(); j++)
{
sbuilder.numLeaves += globalContext->pLocalContext[j]->numLeaves;
sbuilder.numInnerNodes += globalContext->pLocalContext[j]->numInnerNodes;
};
sbuilder.lastInner = globalContext->pLocalContext[numThreads()-1]->lastInnerNode;
sbuilder.lastLeaf = globalContext->pLocalContext[numThreads()-1]->lastLeaf;
// Link the tree
sbuilder.build();
// and run the partitions
LQPartitioner partitioner(localContext);
partitioner.partition();
};
};
// wait for tree to finish
sync();
// now update the copy of the point data
for_loop(pointPartition, LQPointUpdateFunctor(localContext));
// compute the nodes coordinates and sizes
tree.forall_tree_nodes(LQCoordsFunctor(localContext), localContext->innerNodePartition.begin, localContext->innerNodePartition.numNodes)();
tree.forall_tree_nodes(LQCoordsFunctor(localContext), localContext->leafPartition.begin, localContext->leafPartition.numNodes)();
};