// enter a number for a surprise
int main(int argc, char** argv){
if (argc != 2){
puts("That's rather unsportsmanlike. Please include input.");
return 0;
}
int n = atoi(argv[1]);
switch(n){
case 1:
while_loop();
break;
case 4:
for_loop();
break;
case 2:
case 3:
case 7:
case 11:
case 13:
case 17:
case 19:
n_loop(n);
break;
case 1000:
for_loop();
while_loop();
case 6:
case 9:
case 12:
case 15:
case 18:
puts("What strange behavior.");
break;
case 27:
puts("Oh! That's a good number.");
break;
case 100:
puts("Starting countdown...");
for(int i =100; i>0; i--){
if(i == 92){
puts("UGH, this'll take forever.");
continue;
}
printf("%d...\n", i);
}
break;
default:
puts("Nope.");
break;
}
return 0;
}
__AGENCY_ANNOTATION
void for_each(Function&& f)
{
for_loop([&](int i)
{
std::forward<Function>(f)(operator[](i));
});
}
__AGENCY_ANNOTATION
short_vector(Range&& other)
: size_(other.size())
{
assert(other.size() <= max_size());
// copy construct each element with placement new
for_loop([&](int i)
{
T& x = (*this)[i];
::new(&x) T(other[i]);
});
}
void for_loop(qreal *maxSum,
QVector<int> *thresholds,
const QVector<qreal> &H,
int u,
int vmax,
int level,
int levels,
QVector<int> *index)
{
int classes = index->size() - 1;
for (int i = u; i < vmax; i++) {
(*index)[level] = i;
if (level + 1 >= classes) {
// Reached the end of the for loop.
// Calculate the quadratic sum of al intervals.
qreal sum = 0.;
for (int c = 0; c < classes; c++) {
int u = index->at(c);
int v = index->at(c + 1);
sum += H[v + u * levels];
}
if (*maxSum < sum) {
// Return calculated threshold.
*thresholds = index->mid(1, thresholds->size());
*maxSum = sum;
}
} else
// Start a new for loop level, one position after current one.
for_loop(maxSum,
thresholds,
H,
i + 1,
vmax + 1,
level + 1,
levels,
index);
}
}
inline QVector<int> otsu(QVector<int> histogram,
int classes)
{
qreal maxSum = 0.;
QVector<int> thresholds(classes - 1, 0);
QVector<qreal> H = buildTables(histogram);
QVector<int> index(classes + 1);
index[0] = 0;
index[index.size() - 1] = histogram.size() - 1;
for_loop(&maxSum,
&thresholds,
H,
1,
histogram.size() - classes + 1,
1,
histogram.size(), &index);
return thresholds;
}
void FMEMultipoleKernel::multipoleApproxSingleThreaded(ArrayPartition& nodePointPartition)
{
FMELocalContext* localContext = m_pLocalContext;
FMEGlobalContext* globalContext = m_pGlobalContext;
LinearQuadtree& tree = *globalContext->pQuadtree;
if (isMainThread())
{
tree.bottom_up_traversal( // do a bottom up traversal M2M pass
if_then_else(tree.is_leaf_condition(), // if the current node is a leaf
p2m_function(localContext), // then calculate the multipole coeff. due to the points in the leaf
m2m_function(localContext) // else shift the coefficents of all children to center of the inner node
)
)(tree.root());
tree.forall_well_separated_pairs( // do a wspd traversal M2L direct eval
pair_vice_versa(m2l_function(localContext)),// M2L for a well-separated pair
p2p_function(localContext), // direct evaluation
p2p_function(localContext) // direct evaluation
)(tree.root());
tree.top_down_traversal( // top down traversal
if_then_else( tree.is_leaf_condition(), // if the node is a leaf
do_nothing(), // then do nothing, we will deal with this case later
l2l_function(localContext) // else shift the nodes local coeffs to the children
)
)(tree.root());// start at the root
// evaluate all leaves and store the forces in the threads array
for_loop(nodePointPartition, // loop over points
func_comp( // composition of two statements
l2p_function(localContext), // evaluate the forces due to the local expansion in the corresponding leaf
collect_force_function // collect the forces of all threads with the following options:
<
COLLECT_REPULSIVE_FACTOR | // multiply by the repulsive factor stored in the global options
COLLECT_TREE_2_GRAPH_ORDER | // threads data is stored in quadtree leaf order, transform it into graph order
COLLECT_ZERO_THREAD_ARRAY // reset threads array
>(localContext)
)
);
};
};
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)();
};
void FMEMultipoleKernel::operator()(FMEGlobalContext* globalContext)
{
__uint32 maxNumIterations = globalContext->pOptions->maxNumIterations;
__uint32 minNumIterations = globalContext->pOptions->minNumIterations;
__uint32 numPoints = globalContext->pQuadtree->numberOfPoints();
ArrayGraph& graph = *globalContext->pGraph;
LinearQuadtree& tree = *globalContext->pQuadtree;
LinearQuadtreeExpansion& treeExp = *globalContext->pExpansion;
WSPD& wspd = *globalContext->pWSPD;
FMELocalContext* localContext = globalContext->pLocalContext[threadNr()];
FMEGlobalOptions* options = globalContext->pOptions;
float* threadsForceArrayX = localContext->forceX;
float* threadsForceArrayY = localContext->forceY;
float* globalForceArrayX = globalContext->globalForceX;
float* globalForceArrayY = globalContext->globalForceY;
ArrayPartition edgePartition = arrayPartition(graph.numEdges());
ArrayPartition nodePointPartition = arrayPartition(graph.numNodes());
m_pLocalContext = localContext;
m_pGlobalContext = globalContext;
/****************************/
/* INIT */
/****************************/
//! reset the global force array
for_loop_array_set(threadNr(), numThreads(), globalForceArrayX, tree.numberOfPoints(), 0.0f);
for_loop_array_set(threadNr(), numThreads(), globalForceArrayY, tree.numberOfPoints(), 0.0f);
// reset the threads force array
for (__uint32 i = 0; i < tree.numberOfPoints(); i++)
{
threadsForceArrayX[i] = 0.0f;
threadsForceArrayY[i] = 0.0f;
};
__uint32 maxNumIt = options->preProcMaxNumIterations;
for (__uint32 currNumIteration = 0; ((currNumIteration < maxNumIt) ); currNumIteration++)
{
// iterate over all edges and store the resulting forces in the threads array
for_loop(edgePartition,
edge_force_function< EDGE_FORCE_DIV_DEGREE > (localContext) // divide the forces by degree of the node to avoid oscilation
);
// wait until all edges are done
sync();
// now collect the forces in parallel and put the sum into the global array and move the nodes accordingly
for_loop(nodePointPartition,
func_comp(
collect_force_function<COLLECT_EDGE_FACTOR_PREP | COLLECT_ZERO_THREAD_ARRAY >(localContext),
node_move_function<TIME_STEP_PREP | ZERO_GLOBAL_ARRAY>(localContext)
)
);
};
if (isMainThread())
{
globalContext->coolDown = 1.0f;
};
sync();
for (__uint32 currNumIteration = 0; ((currNumIteration < maxNumIterations) && !globalContext->earlyExit); currNumIteration++)
{
// reset the coefficients
for_loop_array_set(threadNr(), numThreads(), treeExp.m_multiExp, treeExp.m_numExp*(treeExp.m_numCoeff << 1), 0.0);
for_loop_array_set(threadNr(), numThreads(), treeExp.m_localExp, treeExp.m_numExp*(treeExp.m_numCoeff << 1), 0.0);
localContext->maxForceSq = 0.0;
localContext->avgForce = 0.0;
// construct the quadtree
quadtreeConstruction(nodePointPartition);
// wait for all threads to finish
sync();
if (isSingleThreaded()) // if is single threaded run the simple approximation
multipoleApproxSingleThreaded(nodePointPartition);
else // otherwise use the partitioning
multipoleApproxFinal(nodePointPartition);
// now wait until all forces are summed up in the global array and mapped to graph node order
sync();
// run the edge forces
for_loop(edgePartition, // iterate over all edges and sum up the forces in the threads array
edge_force_function< EDGE_FORCE_DIV_DEGREE >(localContext) // divide the forces by degree of the node to avoid oscilation
);
// wait until edges are finished
sync();
// collect the edge forces and move nodes without waiting
for_loop(nodePointPartition,
func_comp(
collect_force_function<COLLECT_EDGE_FACTOR | COLLECT_ZERO_THREAD_ARRAY>(localContext),
node_move_function<TIME_STEP_NORMAL | ZERO_GLOBAL_ARRAY>(localContext)
)
);
// wait so we can decide if we need another iteration
sync();
// check the max force square for all threads
if (isMainThread())
{
double maxForceSq = 0.0;
for (__uint32 j=0; j < numThreads(); j++)
maxForceSq = max(globalContext->pLocalContext[j]->maxForceSq, maxForceSq);
// if we are allowed to quit and the max force sq falls under the threshold tell all threads we are done
if ((currNumIteration >= minNumIterations) && (maxForceSq < globalContext->pOptions->stopCritForce ))
{
globalContext->earlyExit = true;
};
};
// this is required to wait for the earlyExit result
sync();
};
};
//! the final approximation algorithm which runs the wspd parallel without storing it in the threads subtrees
void FMEMultipoleKernel::multipoleApproxFinal(ArrayPartition& nodePointPartition)
{
FMELocalContext* localContext = m_pLocalContext;
FMEGlobalContext* globalContext = m_pGlobalContext;
LinearQuadtree& tree = *globalContext->pQuadtree;
// big multihreaded bottom up traversal.
for_tree_partition( // for all roots in the threads tree partition
tree.bottom_up_traversal( // do a bottom up traversal
if_then_else(tree.is_leaf_condition(), // if the current node is a leaf
p2m_function(localContext), // then calculate the multipole coeff. due to the points in the leaf
m2m_function(localContext) // else shift the coefficents of all children to center of the inner node
)
)
);
sync();
// top of the tree has to be done by the main thread
if (isMainThread())
{
tree.bottom_up_traversal( // start a bottom up traversal
if_then_else(tree.is_leaf_condition(), // if the current node is a leaf
p2m_function(localContext), // then calculate the multipole coeff. due to the points in the leaf
m2m_function(localContext) // else shift the coefficents of all children to center of the inner node
),
not_condition(tree.is_fence_condition()))(tree.root());// start at the root, stop when the fence to the threads is reached
tree.forall_well_separated_pairs( // do a wspd traversal
tree.StoreWSPairFunction(), // store the ws pairs in the WSPD
tree.StoreDirectPairFunction(), // store the direct pairs
tree.StoreDirectNodeFunction(), // store the direct nodes
not_condition(tree.is_fence_condition()))(tree.root());
};
// wait for the main thread to finish
sync();
// M2L pass with the WSPD for the result of the single threaded pass above
tree.forall_tree_nodes(M2LFunctor(localContext), localContext->innerNodePartition.begin, localContext->innerNodePartition.numNodes)();
tree.forall_tree_nodes(M2LFunctor(localContext), localContext->leafPartition.begin, localContext->leafPartition.numNodes)();
// D2D pass and store in the thread force array
for_loop(arrayPartition(tree.numberOfDirectPairs()), D2DFunctor(localContext));
for_loop(arrayPartition(tree.numberOfDirectNodes()), NDFunctor(localContext));
// wait until all local coeffs and all direct forces are computed
sync();
// the rest of the WSPD can be done on the fly by the thread
for_tree_partition(
tree.forall_well_separated_pairs( // do a wspd traversal
pair_vice_versa(m2l_function(localContext)), // M2L for a well-separated pair
p2p_function(localContext), // direct evaluation
p2p_function(localContext) // direct evaluation
)
);
// wait until all local coeffs and all direct forces are computed
sync();
// big multihreaded top down traversal. top of the tree has to be done by the main thread
if (isMainThread())
{
tree.top_down_traversal( // top down traversal L2L pass
if_then_else( tree.is_leaf_condition(), // if the node is a leaf
do_nothing(), // then do nothing, we will deal with this case later
l2l_function(localContext) // else shift the nodes local coeffs to the children
),
not_condition(tree.is_fence_condition()) // stop when the fence to the threads is reached
)(tree.root()); // start at the root,
};
// wait for the top of the tree
sync();
for_tree_partition( // for all roots in the threads tree partition L2L pass
tree.top_down_traversal( // do a top down traversal
if_then_else( tree.is_leaf_condition(), // if the node is a leaf
do_nothing(), // then do nothing, we will deal with this case later
l2l_function(localContext) // else shift the nodes local coeffs to the children
)
)
);
// wait until the traversal is finished and all leaves have their accumulated local coeffs
sync();
// evaluate all leaves and store the forces in the threads array (Note we can store them in the global array but then we have to use random access)
// we can start immediately to collect the forces because we evaluated before point by point
for_loop(nodePointPartition, // loop over threads points
func_comp( // composition of two statements
l2p_function(localContext), // evaluate the forces due to the local expansion in the corresponding leaf
collect_force_function // collect the forces of all threads with the following options:
<
COLLECT_REPULSIVE_FACTOR | // multiply by the repulsive factor stored in the global options
COLLECT_TREE_2_GRAPH_ORDER | // threads data is stored in quadtree leaf order, transform it into graph order
COLLECT_ZERO_THREAD_ARRAY // reset threads array
>(localContext)
)
);
};
struct Value execute (struct Tree * ast, struct Tree_map * defined, struct Map * let_map){
struct Value result;
// first check for special kinds of execution
if(ast->type == 'k' && string_matches(&ast->content.data.str, &if_const)){
return if_block(ast, defined, let_map);
}
else if(ast->type == 'k' && string_matches(&let_const, &ast->content.data.str)){
store_let_binding(ast, defined, let_map);
result.type = 'u';
}
else if(ast->type == 'k' && string_matches(&each_const, &ast->content.data.str)){
for_each(ast, defined, let_map);
result.type = 'u';
}
else if(ast->type == 'k' && string_matches(&map_const, &ast->content.data.str)){
return map_array(ast, defined, let_map);
}
else if(ast->type == 'k' && string_matches(&reduce_const, &ast->content.data.str)){
return reduce_array(ast, defined, let_map);
}
else if(ast->type == 'k' && string_matches(&set_const, &ast->content.data.str)){
struct Value index = execute(ast->children[0], defined, let_map);
struct Value item = execute(ast->children[1], defined, let_map);
struct Value array = execute(ast->children[2], defined, let_map);
result = array_set(index, item, array);
return result;
}
else if(ast->type == 'k' && string_matches(&for_const, &ast->content.data.str)){
for_loop(ast, defined, let_map);
result.type = 'u'; //return undefined
}
else if(ast->type == 'k' && string_matches(&do_const, &ast->content.data.str)){
for(int i = 0; i < ast->size; i++){
if(i == ast->size-1){
result = execute(ast->children[i], defined, let_map);
} else {
execute(ast->children[i], defined, let_map);
}
}
}
else if(ast->type == 'k' && string_matches(&read_const, &ast->content.data.str)){
return read_file(ast->children[0]->content.data.str);
}
else if(ast->type == 'k' && string_matches(&substring_const, &ast->content.data.str)){
struct Value string = execute(ast->children[2], defined, let_map);
struct Value start = execute(ast->children[0], defined, let_map);
struct Value end = execute(ast->children[1], defined, let_map);
if(string.type != 's'){
ERROR("Non-string value passed into substring: %c.", string.type);
result.type = 'u';
return result;
} else {
return substring(start.data.ln, end.data.ln, string);
}
}
else if(ast->type == 'k' && string_matches(&switch_const, &ast->content.data.str)){
return switch_case(ast, defined, let_map);
}
else if(ast->type == 'k' && string_matches(¶llel_const, &ast->content.data.str)){
parallel_execution(ast, defined, let_map);
result.type = 'u';
} else {
// no special execution types found, check for more basic conditions
int idx;
if(!ast->size){
// ast with no children is either a value or a variable
if(ast->type == 'k'){
for(int i = 0; i < let_map->size; i++){
if(string_matches(&let_map->members[i]->key->data.str, &ast->content.data.str)){
return *let_map->members[i]->val;
}
}
ERROR("Undefined variable: %s", ast->content.data.str.body);
} else {
return ast->content;
}
}
else if(ast->type == 'k' && (idx = is_defined_func(defined, ast->content.data.str)) > -1){
return execute_defined_func(ast, defined, let_map, idx);
}
else if(ast->size == 1){
struct Value a = execute(ast->children[0], defined, let_map);
if(ast->type == 'k'){
if(string_matches(&ast->content.data.str, &print_const)){
print(a);
printf("\n");
result.type = 'u';
}
else if(string_matches(&ast->content.data.str, &length_const)){
return length(a);
}
else if(string_matches(&ast->content.data.str, &return_const)){
return execute(ast->children[0], defined, let_map);
}
}
}
else if(ast->size == 2) {
struct Value a = execute(ast->children[0], defined, let_map);
struct Value b = execute(ast->children[1], defined, let_map);
result = apply_core_function(ast, a, b);
} else {
result = reduce_ast(ast, defined, let_map);
}
}
return result;
}