Commit 71c48282 authored by Giovanni La Mura's avatar Giovanni La Mura
Browse files

Use pre-computed indices in loop and fix timers

parent 5795b9a1

src/trapping/cfrfme.cpp

+15 −13
Original line number Diff line number Diff line
@@ -279,7 +279,7 @@ void frfme(string data_file, string output_path) { 
#ifdef USE_NVTX

    nvtxRangePop();

#endif

  } else { // label 16; jlfm = 1

  } else { // label 16; jlmf = 1

#ifdef USE_NVTX

    nvtxRangePush("frfme() with jlmf == 1");

#endif
@@ -539,7 +539,8 @@ void frfme(string data_file, string output_path) { 
	  double *vec_vkzm = vkzm[0];

	  dcomplex *global_vec_w = new dcomplex[size_global_vec_w];

#ifdef USE_TARGET_OFFLOAD

	  frfme_duration = t_start - chrono::high_resolution_clock::now();

	  elapsed = t_start - chrono::high_resolution_clock::now();

	  frfme_duration = elapsed;

	  t_start = chrono::high_resolution_clock::now();

	  message = "INFO: Mapping data to device.\n";

	  logger.log(message);
@@ -706,11 +707,12 @@ void frfme(string data_file, string output_path) { 
#ifdef USE_NVTX

  nvtxRangePop();

#endif

#ifndef USE_TARGET_OFFLOAD

  elapsed = chrono::high_resolution_clock::now() - t_start;

#ifdef USE_TARGET_OFFLOAD

  frfme_duration += elapsed;

#else

  frfme_duration = elapsed;

#else

  elapsed = chrono::high_resolution_clock::now() - t_end;

  frfme_duration += elapsed;

#endif

  message = "INFO: FRFME took " + to_string(frfme_duration.count()) + "s.\n";

  logger.log(message);
@@ -753,10 +755,10 @@ void offload_loop( 
  double *_yv, const int nyv, double *_zv, const int nzv, double *vec_vkzm, const int jlmf, const int jlml,

  const int nkv, const int nlmmt, double delks, double frsh

) {

  int nrvc = nxv * nyv * nzv;

  int nvtot = nxv * nyv * nzv;

  int nkvs = nkv * nkv;

  int nkvmo = nkv - 1;

  int nkvvmo = (nkv - 1) * nkv;

  int nkvmo = nkvmo;

  int nkvvmo = nkvmo * nkv;

  int nvxy = nxv * nyv;

  dcomplex cc0 = 0.0 + I * 0.0;

  dcomplex uim = 0.0 + I * 1.0;
@@ -764,7 +766,7 @@ void offload_loop( 
  for (int j80 = jlmf - 1; j80 < jlml; j80++) {

    int j80_index = j80 - jlmf + 1;

    dcomplex *vec_w = global_vec_w + nkvs * j80_index;

    for (int ixyz = 0; ixyz < nrvc; ixyz++) {

    for (int ixyz = 0; ixyz < nvtot; ixyz++) {

      int iz75 = ixyz / nvxy;

      int iy70 = (ixyz % nvxy) / nxv;

      int ix65 = ixyz % nxv;
@@ -779,7 +781,7 @@ void offload_loop( 
	int jx55 = jy60x55 % nkv;

	int w_index = (jx55 * nkv) + jy60;

	double vky = vkv[jy60];

	double factor = (jy60 == 0 || jy60 == nkv - 1) ? 0.5 : 1.0;

	double factor = (jy60 == 0 || jy60 == nkvmo) ? 0.5 : 1.0;

	if (jx55 == 0) {

	  // jx55 = 0: phasf

	  double vkx = vkv[nkvmo];
@@ -790,9 +792,9 @@ void offload_loop( 
	  term *= factor;

	  rsumy += (real(term));

	  isumy += (imag(term));

	} else if (jx55 == nkv - 1) {

	} else if (jx55 == nkvmo) {

	  // jx55 = nkv - 1: phasl

	  double vkx = vkv[nkv - 1];

	  double vkx = vkv[nkvmo];

	  double vkz = vec_vkzm[nkvvmo + jy60];

	  double angle = vkx * x + vky * y + vkz * z;

	  dcomplex phasl = cos(angle) + uim * sin(angle);
@@ -813,7 +815,7 @@ void offload_loop( 
	}

      } // jy60x55 loop

      dcomplex sumy = rsumy + uim * isumy;

      vec_wsum[(j80_index * nrvc) + ixyz] = sumy * delks;

      vec_wsum[(j80_index * nvtot) + ixyz] = sumy * delks;

    } // ixyz loop

  } // j80 loop

}