PDS_TriangleCount/src/v4.cpp

/*!
 * \file    v4.cpp
 * \brief   vv3 part of the exercise.
 *
 * \author
 *    Christos Choutouridis AEM:8997
 *    <cchoutou@ece.auth.gr>
 */
#include <v4.h>

namespace v4 {

#if defined CILK

// export CILK_NWORKERS=<num>
int nworkers() {
   if (session.max_threads)
      return (session.max_threads < __cilkrts_get_nworkers()) ?
            session.max_threads : __cilkrts_get_nworkers();
   else
      return __cilkrts_get_nworkers();
}

std::vector<value_t> mmacc_v(matrix& A, matrix& B) {
   std::vector<value_t> c(A.size());

   cilk_for (int i=0 ; i<A.size() ; ++i) {
      for (auto j = A.getRow(i); j.index() != j.end() ; ++j){
         c[i] += A.getRow(i)*B.getCol(j.index());
      }
   }
   if (session.makeSymmetric)
      std::transform (c.begin(), c.end(), c.begin(), [] (value_t& x) {
         return x/2;
      });
   return c;
}

void do_sum (value_t& out_sum, std::vector<value_t>& v, index_t begin, index_t end) {
   for (auto i =begin ; i != end ; ++i)
      out_sum += v[i];
}

value_t sum (std::vector<value_t>& v) {
   int n = nworkers();
   std::vector<value_t> sum_v(n, 0);

   for (index_t i =0 ; i < n ; ++i) {
      cilk_spawn do_sum(sum_v[i], v, i*v.size()/n, (i+1)*v.size()/n);
   }
   cilk_sync;

   value_t s =0;
   for (auto& it : sum_v) s += it;
   return s;
}

#elif defined OMP

/*
// export OMP_NUM_THREADS=<num>
 */
int nworkers() {
   if (session.max_threads && session.max_threads < (size_t)omp_get_max_threads()) {
      omp_set_dynamic(0);
      omp_set_num_threads(session.max_threads);
      return session.max_threads;
   }
   else {
      omp_set_dynamic(1);
      return omp_get_max_threads();
   }
}

std::vector<value_t> mmacc_v(matrix& A, matrix& B) {
   std::vector<value_t> c(A.size());

   #pragma omp parallel for shared(c)
   for (int i=0 ; i<A.size() ; ++i) {
      for (auto j = A.getRow(i); j.index() != j.end() ; ++j) {
         c[i] += A.getRow(i)*B.getCol(j.index());
      }
   }
   if (session.makeSymmetric)
      std::transform (c.begin(), c.end(), c.begin(), [] (value_t& x) {
         return x/2;
      });
   return c;
}

value_t sum (std::vector<value_t>& v) {
   value_t s =0;

   #pragma omp parallel for reduction(+:s)
   for (auto i =0u ; i<v.size() ; ++i)
      s += v[i];
   return s;
}

#elif defined THREADS

/*
 * std::thread::hardware_concurrency()
 */
int nworkers() {
   if (session.max_threads)
      return (session.max_threads < std::thread::hardware_concurrency()) ?
            session.max_threads : std::thread::hardware_concurrency();
   else
      return std::thread::hardware_concurrency();
}

std::vector<value_t> mmacc_v_rng(std::vector<value_t>& out, matrix& A, matrix& B, index_t begin, index_t end) {
   for (index_t i=begin ; i<end ; ++i) {
      for (auto j = A.getRow(i); j.index() != j.end() ; ++j){
         out[i] += A.getRow(i)*B.getCol(j.index());
      }
   }
   return out;
}

std::vector<value_t> mmacc_v(matrix& A, matrix& B) {
   std::vector<std::thread> workers;
   std::vector<value_t> c(A.size());
   int n = nworkers();

   for (index_t i=0 ; i<n ; ++i)
      workers.push_back (std::thread (mmacc_v_rng, std::ref(c), std::ref(A), std::ref(B), i*c.size()/n, (i+1)*c.size()/n));

   std::for_each(workers.begin(), workers.end(), [](std::thread& t){
      t.join();
   });
   if (session.makeSymmetric)
      std::transform (c.begin(), c.end(), c.begin(), [] (value_t& x) {
         return x/2;
      });
   return c;
}

void do_sum (value_t& out_sum, std::vector<value_t>& v, index_t begin, index_t end) {
   for (auto i =begin ; i != end ; ++i)
      out_sum += v[i];
}

value_t sum (std::vector<value_t>& v) {
   int n = nworkers();
   std::vector<value_t> sum_v(n, 0);
   std::vector<std::thread> workers;

   for (index_t i =0 ; i < n ; ++i)
      workers.push_back (std::thread (do_sum, std::ref(sum_v[i]), std::ref(v), i*v.size()/n, (i+1)*v.size()/n));

   std::for_each(workers.begin(), workers.end(), [](std::thread& t){
      t.join();
   });

   value_t s =0;
   for (auto& it : sum_v) s += it;
   return s;
}

#else

int nworkers() { return 1; }

std::vector<value_t> mmacc_v(matrix& A, matrix& B) {
   std::vector<value_t> c(A.size());
   for (int i=0 ; i<A.size() ; ++i) {
      for (auto j = A.getRow(i); j.index() != j.end() ; ++j){
         c[i] += A.getRow(i)*B.getCol(j.index());
      }
   }
   if (session.makeSymmetric)
      std::transform (c.begin(), c.end(), c.begin(), [] (value_t& x) {
         return x/2;
      });
   return c;
}

value_t sum (std::vector<value_t>& v) {
   value_t s =0;
   for (auto& it : v)
      s += it;
   return s;
}

#endif

std::vector<value_t> triang_v(matrix& A) {
   return mmacc_v(A, A);
}

value_t triang_count (std::vector<value_t>& c) {
   return (session.makeSymmetric) ? sum(c)/3 : sum(c);
}

}  // namespace v4