Commit 9d8d0024 authored by Giovanni La Mura's avatar Giovanni La Mura
Browse files

Initialize thread-local outputs only when actually needed

parent 2028cb6f

src/cluster/cluster.cpp

+11 −6
Original line number Diff line number Diff line
@@ -404,6 +404,7 @@ void cluster(const string& config_file, const string& data_file, const string& o 
	  cid_2 = cid;

	  // OMP thread 0 of MPI process 0 holds the pointer to the full output structure

	  p_output_2 = p_output;

	  p_outarray[0] = p_output_2;

	} else {

	  // this is not thread 0, so do create fresh copies of all local variables

	  cid_2 = new ClusterIterationData(*cid);
@@ -422,7 +423,6 @@ void cluster(const string& config_file, const string& data_file, const string& o 
	  vtppoanp_2 = new VirtualBinaryFile();

	  // each thread opens new virtual files and stores their pointers in the shared array

	  // each thread puts a copy of the pointers to its virtual files in the shared arrays

	  p_outarray[myompthread] = p_output_2;

	  vtppoanarray[myompthread] = vtppoanp_2;

#pragma omp barrier
@@ -431,6 +431,7 @@ void cluster(const string& config_file, const string& data_file, const string& o 
	    if (myompthread > 0) {

	      // UPDATE: non-0 threads need to allocate memory for one scale at a time.

	      p_output_2 = new ClusterOutputInfo(sconf, gconf, mpidata, myjxi488, 1);

	      p_outarray[myompthread] = p_output_2;

	    }

	    int jer = cluster_jxi488_cycle(myjxi488, sconf, gconf, p_scattering_angles, cid_2, p_output_2, output_path, vtppoanp_2);

	  }
@@ -446,6 +447,7 @@ void cluster(const string& config_file, const string& data_file, const string& o 
	      if (p_outarray[ti] != NULL) {

		p_outarray[0]->insert(*(p_outarray[ti]));

		delete p_outarray[ti];

		p_outarray[ti] = NULL;

	      }

	      vtppoanarray[0]->append(*(vtppoanarray[ti]));

	      delete vtppoanarray[ti];
@@ -594,14 +596,9 @@ void cluster(const string& config_file, const string& data_file, const string& o 
	// the parallel loop over MPI processes covers a different set of indices for each thread

#pragma omp barrier

	int myjxi488 = ixi488 + myjxi488startoffset + myompthread;

	// Thread 0 of non-zero MPI processes needs to allocate memory for the

	// output of all threads.

	if (myompthread == 0)

	  p_output_2 = new ClusterOutputInfo(sconf, gconf, mpidata, myjxi488, ompnumthreads);

	// each thread opens new virtual files and stores their pointers in the shared array

	vtppoanp_2 = new VirtualBinaryFile();

	// each thread puts a copy of the pointers to its virtual files in the shared arrays

	p_outarray[myompthread] = p_output_2;

	vtppoanarray[myompthread] = vtppoanp_2;

#pragma omp barrier

	if (myompthread==0) logger->log("Syncing OpenMP threads and starting the loop on wavelengths\n");
@@ -609,7 +606,14 @@ void cluster(const string& config_file, const string& data_file, const string& o 
	// each MPI process handles a number of contiguous scales corresponding to its number of OMP threads at this omp level of parallelism

	if (myjxi488 <= cid_2->number_of_scales) {

	  if (myompthread > 0) {

	    // UPDATE: non-0 threads need to allocate memory for one scale at a time.

	    p_output_2 = new ClusterOutputInfo(sconf, gconf, mpidata, myjxi488, 1);

	    p_outarray[myompthread] = p_output_2;

	  } else {

	    // Thread 0 of non-zero MPI processes needs to allocate memory for the

	    // output of all threads.

	    p_output_2 = new ClusterOutputInfo(sconf, gconf, mpidata, myjxi488, ompnumthreads);

	    p_outarray[0] = p_output_2;

	  }

	  int jer = cluster_jxi488_cycle(myjxi488, sconf, gconf, p_scattering_angles, cid_2, p_output_2, output_path, vtppoanp_2);

	} // close the OMP parallel for loop
@@ -621,6 +625,7 @@ void cluster(const string& config_file, const string& data_file, const string& o 
	    if (p_outarray[ti] != NULL) {

	      p_outarray[0]->insert(*(p_outarray[ti]));

	      delete p_outarray[ti];

	      p_outarray[ti] = NULL;

	    }

	    vtppoanarray[0]->append(*(vtppoanarray[ti]));

	    delete vtppoanarray[ti];