Commit abea8076 authored by Giovanni La Mura's avatar Giovanni La Mura
Browse files

Update OpenMP implementation of frfme()

parent 0b7e3963

src/trapping/cfrfme.cpp

+21 −20
Original line number Diff line number Diff line
@@ -129,19 +129,19 @@ void frfme(string data_file, string output_path) { 
      if (tt2 != NULL) {

	vkv = tt2->get_vector();

	vkzm = tt2->get_matrix();

	apfafa = tt2->get_param("apfafa");

	pmf = tt2->get_param("pmf");

	spd = tt2->get_param("spd");

	rir = tt2->get_param("rir");

	ftcn = tt2->get_param("ftcn");

	fshmx = tt2->get_param("fshmx");

	vxyzmx = tt2->get_param("vxyzmx");

	delxyz = tt2->get_param("delxyz");

	vknmx = tt2->get_param("vknmx");

	delk = tt2->get_param("delk");

	delks = tt2->get_param("delks");

	nlmmt = (int)tt2->get_param("nlmmt");

	nrvc = (int)tt2->get_param("nrvc");

	apfafa = tt2->apfafa;

	pmf = tt2->pmf;

	spd = tt2->spd;

	rir = tt2->rir;

	ftcn = tt2->ftcn;

	fshmx = tt2->fshmx;

	vxyzmx = tt2->vxyzmx;

	delxyz = tt2->delxyz;

	vknmx = tt2->vknmx;

	delk = tt2->delk;

	delks = tt2->delks;

	nlmmt = tt2->nlmmt;

	nrvc = tt2->nrvc;

      } else {

	printf("ERROR: could not open TEMPTAPE2 file.\n");

      }
@@ -273,10 +273,10 @@ void frfme(string data_file, string output_path) { 
	long swap1_size, swap2_size, tfrfme_size;

	double size_mb;

	printf("INFO: calculating memory requirements\n");

	swap1_size = Swap1::get_memory_requirement(lm, nkv);

	swap1_size = Swap1::get_size(lm, nkv);

	size_mb = 1.0 * swap1_size / 1024.0 / 1024.0;

	printf("Swap 1: %.2lg MB\n", size_mb);

	swap2_size = Swap2::get_memory_requirement(nkv);

	swap2_size = Swap2::get_size(nkv);

	size_mb = 1.0 * swap2_size / 1024.0 / 1024.0;

	printf("Swap 2: %.2lg MB\n", size_mb);

	tt2 = new Swap2(nkv);
@@ -374,13 +374,14 @@ void frfme(string data_file, string output_path) { 
#ifdef USE_NVTX

	  nvtxRangePop();

#endif

	  dcomplex *vec_w = new dcomplex[nkv * nkv]();

	  dcomplex **w = new dcomplex*[nkv];

	  for (int wi = 0; wi < nkv; wi++) w[wi] = vec_w + wi * nkv;

#ifdef USE_NVTX

	  nvtxRangePush("j80 loop");

#endif

#pragma omp parallel for

	  for (int j80 = jlmf; j80 <= jlml; j80++) {

	    dcomplex *vec_w = new dcomplex[nkv * nkv]();

	    dcomplex **w = new dcomplex*[nkv];

	    for (int wi = 0; wi < nkv; wi++) w[wi] = vec_w + wi * nkv;

	    int wk_index = 0;

	    for (int jy50 = 0; jy50 < nkv; jy50++) {

	      for (int jx50 = 0; jx50 < nkv; jx50++) {
@@ -419,9 +420,9 @@ void frfme(string data_file, string output_path) { 
		} // ix65 loop

	      } // iy70 loop

	    } // iz75 loop

	  } // j80 loop

	    delete[] vec_w;

	    delete[] w;

	  } // j80 loop

#ifdef USE_NVTX

	  nvtxRangePop();

#endif