add part_load col

src/sync/sync.h

@@ -21,6 +21,9 @@ enum ReduceOp {
 
 /*! \brief get rank of current process */
 int GetRank(void);
+/*! \brief get total number of process */
+int GetWorldSize(void);
+
 /*! 
  * \brief this is used to check if sync module is a true distributed implementation, or simply a dummpy
  */

src/sync/sync_empty.cpp

@@ -17,6 +17,10 @@ bool IsDistributed(void) {
   return false;
 }
 
+int GetWorldSize(void) {
+  return 1;
+}
+
 template<>
 void AllReduce<uint32_t>(uint32_t *sendrecvbuf, int count, ReduceOp op) {
 }

src/sync/sync_mpi.cpp

@@ -8,6 +8,10 @@ int GetRank(void) {
   return MPI::COMM_WORLD.Get_rank();
 }
 
+int GetWorldSize(void) {
+  return MPI::COMM_WORLD.Get_size();
+}
+
 void Init(int argc, char *argv[]) {
   MPI::Init(argc, argv);
 }