void parallel_dc(graph_t& g, unsigned threadnum, gBenchPerf_multi & perf, int perf_group)
{
uint64_t chunk = (unsigned)ceil(g.num_vertices()/(double)threadnum);
#pragma omp parallel num_threads(threadnum)
{
unsigned tid = omp_get_thread_num();
perf.open(tid, perf_group);
perf.start(tid, perf_group);
unsigned start = tid*chunk;
unsigned end = start + chunk;
if (end > g.num_vertices()) end = g.num_vertices();
for (unsigned vid=start;vid<end;vid++)
{
vertex_iterator vit = g.find_vertex(vid);
// out degree
vit->property().outdegree = vit->edges_size();
// in degree
edge_iterator eit;
for (eit=vit->edges_begin(); eit!=vit->edges_end(); eit++)
{
vertex_iterator targ = g.find_vertex(eit->target());
__sync_fetch_and_add(&(targ->property().indegree), 1);
}
}
perf.stop(tid, perf_group);
}
}
size_t parallel_kcore(graph_t& g, size_t k, unsigned threadnum, gBenchPerf_multi & perf, int perf_group)
{
// initialize
size_t remove_cnt=0;
vector<vector<uint64_t> > global_input_tasks(threadnum);
vector<vector<uint64_t> > global_output_tasks(threadnum*threadnum);
for (vertex_iterator vit=g.vertices_begin(); vit!=g.vertices_end(); vit++)
{
size_t degree = vit->edges_size();
vit->property().degree = degree;
if (degree < k)
{
global_input_tasks[vertex_distributor(vit->id(), threadnum)].push_back(vit->id());
vit->property().removed = true;
remove_cnt++;
}
}
bool stop = false;
#pragma omp parallel num_threads(threadnum) shared(stop,global_input_tasks,global_output_tasks)
{
unsigned tid = omp_get_thread_num();
vector<uint64_t> & input_tasks = global_input_tasks[tid];
perf.open(tid, perf_group);
perf.start(tid, perf_group);
while(!stop)
{
#pragma omp barrier
// process local queue
stop = true;
for (unsigned i=0;i<input_tasks.size();i++)
{
uint64_t vid=input_tasks[i];
vertex_iterator vit = g.find_vertex(vid);
for (edge_iterator eit=vit->edges_begin();eit!=vit->edges_end();eit++)
{
uint64_t dest_vid = eit->target();
vertex_iterator destvit = g.find_vertex(dest_vid);
if (destvit->property().removed==false)
{
unsigned oldval = __sync_fetch_and_sub(&(destvit->property().degree), 1);
if (oldval == k)
{
destvit->property().removed=true;
__sync_fetch_and_add(&remove_cnt, 1);
global_output_tasks[vertex_distributor(dest_vid,threadnum)+tid*threadnum].push_back(dest_vid);
}
}
}
}
#pragma omp barrier
input_tasks.clear();
for (unsigned i=0;i<threadnum;i++)
{
if (global_output_tasks[i*threadnum+tid].size()!=0)
{
stop = false;
input_tasks.insert(input_tasks.end(),
global_output_tasks[i*threadnum+tid].begin(),
global_output_tasks[i*threadnum+tid].end());
global_output_tasks[i*threadnum+tid].clear();
}
}
#pragma omp barrier
}
perf.stop(tid, perf_group);
}
return remove_cnt;
}
void parallel_wcc(graph_t &g, unsigned threadnum, vector<vector<uint64_t> > &global_input_tasks, gBenchPerf_multi &perf,
int perf_group)
{
vector<vector<uint64_t> > global_output_tasks(threadnum*threadnum);
bool stop = false;
#pragma omp parallel num_threads(threadnum) shared(stop,global_input_tasks,global_output_tasks,perf)
{
unsigned tid = omp_get_thread_num();
vector<uint64_t> & input_tasks = global_input_tasks[tid];
perf.open(tid, perf_group);
perf.start(tid, perf_group);
while(!stop)
{
#pragma omp barrier
// process local queue
stop = true;
for (unsigned i=0;i<input_tasks.size();i++)
{
uint64_t vid=input_tasks[i];
uint64_t size, begin;
size = g.csr_in_edges_size(vid);
begin = g.csr_in_edges_begin(vid);
for (uint64_t i=0;i<size;i++)
{
uint64_t dest_vid = g.csr_in_edge(begin,i);
if(g.csr_vertex_property(dest_vid).root > g.csr_vertex_property(vid).root) {
__sync_bool_compare_and_swap(&(g.csr_vertex_property(dest_vid).root), g.csr_vertex_property(dest_vid).root, g.csr_vertex_property(vid).root);
global_output_tasks[vertex_distributor(dest_vid,threadnum)+tid*threadnum].push_back(dest_vid);
}
}
size = g.csr_out_edges_size(vid);
begin = g.csr_out_edges_begin(vid);
for (uint64_t i=0;i<size;i++)
{
uint64_t dest_vid = g.csr_out_edge(begin,i);
bool done = false;
while(!done) {
if(g.csr_vertex_property(dest_vid).root > g.csr_vertex_property(vid).root) {
done = __sync_bool_compare_and_swap(&(g.csr_vertex_property(dest_vid).root), g.csr_vertex_property(dest_vid).root, g.csr_vertex_property(vid).root);
if(done) {
global_output_tasks[vertex_distributor(dest_vid,threadnum)+tid*threadnum].push_back(dest_vid);
}
} else {
done = true;
}
}
}
}
#pragma omp barrier
input_tasks.clear();
for (unsigned i=0;i<threadnum;i++)
{
if (global_output_tasks[i*threadnum+tid].size()!=0)
{
stop = false;
input_tasks.insert(input_tasks.end(),
global_output_tasks[i*threadnum+tid].begin(),
global_output_tasks[i*threadnum+tid].end());
global_output_tasks[i*threadnum+tid].clear();
}
}
#pragma omp barrier
}
perf.stop(tid, perf_group);
}
}
void parallel_bc(graph_t& g, unsigned threadnum, bool undirected,
gBenchPerf_multi & perf, int perf_group)
{
typedef list<size_t> vertex_list_t;
size_t vnum = g.num_vertices();
uint64_t chunk = (unsigned)ceil(vnum/(double)threadnum);
double normalizer;
normalizer = (undirected)? 2.0 : 1.0;
#pragma omp parallel num_threads(threadnum)
{
unsigned tid = omp_get_thread_num();
perf.open(tid, perf_group);
perf.start(tid, perf_group);
unsigned start = tid*chunk;
unsigned end = start + chunk;
if (maxiter != 0 && chunk > maxiter) end = start + maxiter;
if (end > vnum) end = vnum;
// initialization
vector<vertex_list_t> shortest_path_parents(vnum);
vector<int16_t> num_of_paths(vnum);
vector<uint16_t> depth_of_vertices(vnum); // 16 bits signed
vector<float> centrality_update(vnum);
#ifdef SIM
unsigned iter = 0;
#endif
for (uint64_t vid=start;vid<end;vid++)
{
#ifdef SIM
SIM_BEGIN(iter==beginiter);
iter++;
#endif
size_t vertex_s = vid;
stack<size_t> order_seen_stack;
queue<size_t> BFS_queue;
BFS_queue.push(vertex_s);
for (size_t i=0;i<vnum;i++)
{
shortest_path_parents[i].clear();
num_of_paths[i] = (i==vertex_s) ? 1 : 0;
depth_of_vertices[i] = (i==vertex_s) ? 0: MY_INFINITY;
centrality_update[i] = 0;
}
// BFS traversal
while (!BFS_queue.empty())
{
size_t v = BFS_queue.front();
BFS_queue.pop();
order_seen_stack.push(v);
vertex_iterator vit = g.find_vertex(v);
uint16_t newdepth = depth_of_vertices[v]+1;
for (edge_iterator eit=vit->edges_begin(); eit!= vit->edges_end(); eit++)
{
size_t w = eit->target();
#ifdef HMC
if (HMC_CAS_equal_16B(&(depth_of_vertices[w]),MY_INFINITY,newdepth) == MY_INFINITY)
{
BFS_queue.push(w);
}
if (depth_of_vertices[w] == newdepth)
{
HMC_ADD_16B(&(num_of_paths[w]), num_of_paths[v]);
shortest_path_parents[w].push_back(v);
}
#else
if (depth_of_vertices[w] == MY_INFINITY)
{
BFS_queue.push(w);
depth_of_vertices[w] = newdepth;
}
if (depth_of_vertices[w] == newdepth)
{
num_of_paths[w] += num_of_paths[v];
shortest_path_parents[w].push_back(v);
}
#endif
}
}
// dependency accumulation
while (!order_seen_stack.empty())
{
size_t w = order_seen_stack.top();
order_seen_stack.pop();
float coeff = (1+centrality_update[w])/(double)num_of_paths[w];
vertex_list_t::iterator iter;
for (iter=shortest_path_parents[w].begin();
iter!=shortest_path_parents[w].end(); iter++)
{
size_t v=*iter;
#ifdef HMC
HMC_FP_ADD(&(centrality_update[v]), (num_of_paths[v]*coeff));
#else
centrality_update[v] += (num_of_paths[v]*coeff);
#endif
}
if (w!=vertex_s)
{
vertex_iterator vit = g.find_vertex(w);
#pragma omp atomic
vit->property().BC += centrality_update[w]/normalizer;
}
}
#ifdef SIM
SIM_END(iter==enditer);
#endif
}
#ifdef SIM
SIM_END(enditer==0);
#endif
perf.stop(tid, perf_group);
}
return;
}
void parallel_bc(graph_t& g, unsigned threadnum, bool undirected,
gBenchPerf_multi & perf, int perf_group)
{
typedef list<size_t> vertex_list_t;
size_t vnum = g.num_vertices();
uint64_t chunk = (unsigned)ceil(vnum/(double)threadnum);
double normalizer;
normalizer = (undirected)? 2.0 : 1.0;
#pragma omp parallel num_threads(threadnum)
{
unsigned tid = omp_get_thread_num();
perf.open(tid, perf_group);
perf.start(tid, perf_group);
unsigned start = tid*chunk;
unsigned end = start + chunk;
if (end > vnum) end = vnum;
// initialization
vector<vertex_list_t> shortest_path_parents(vnum);
vector<size_t> num_of_paths(vnum);
vector<int8_t> depth_of_vertices(vnum); // 8 bits signed
vector<double> centrality_update(vnum);
for (uint64_t vid=start;vid<end;vid++)
{
size_t vertex_s = vid;
stack<size_t> order_seen_stack;
queue<size_t> BFS_queue;
BFS_queue.push(vertex_s);
for (size_t i=0;i<vnum;i++)
{
shortest_path_parents[i].clear();
num_of_paths[i] = (i==vertex_s) ? 1 : 0;
depth_of_vertices[i] = (i==vertex_s) ? 0: -1;
centrality_update[i] = 0;
}
// BFS traversal
while (!BFS_queue.empty())
{
size_t v = BFS_queue.front();
BFS_queue.pop();
order_seen_stack.push(v);
vertex_iterator vit = g.find_vertex(v);
for (edge_iterator eit=vit->edges_begin(); eit!= vit->edges_end(); eit++)
{
size_t w = eit->target();
if (depth_of_vertices[w]<0)
{
BFS_queue.push(w);
depth_of_vertices[w] = depth_of_vertices[v] + 1;
}
if (depth_of_vertices[w] == (depth_of_vertices[v] + 1))
{
num_of_paths[w] += num_of_paths[v];
shortest_path_parents[w].push_back(v);
}
}
}
// dependency accumulation
while (!order_seen_stack.empty())
{
size_t w = order_seen_stack.top();
order_seen_stack.pop();
double coeff = (1+centrality_update[w])/(double)num_of_paths[w];
vertex_list_t::iterator iter;
for (iter=shortest_path_parents[w].begin();
iter!=shortest_path_parents[w].end(); iter++)
{
size_t v=*iter;
centrality_update[v] += (num_of_paths[v]*coeff);
}
if (w!=vertex_s)
{
vertex_iterator vit = g.find_vertex(w);
#pragma omp atomic
vit->property().BC += centrality_update[w]/normalizer;
}
}
}
perf.stop(tid, perf_group);
}
return;
}
