Commit 4dfc33af authored by Giovanni La Mura's avatar Giovanni La Mura
Browse files

Use aligned vector allocations when offloading to GPU

parent 57ac3b1a

src/libnptm/tfrfme.cpp

+39 −4
Original line number Diff line number Diff line
@@ -44,6 +44,10 @@ 
#include "../include/file_io.h"

#endif



#ifdef USE_TARGET_OFFLOAD

#include <cstdlib>

#endif



using namespace std;



// >>> START OF Swap1 CLASS IMPLEMENTATION <<<
@@ -220,15 +224,32 @@ bool Swap1::operator ==(Swap1 &other) { 
// >>> START OF Swap2 CLASS IMPLEMENTATION <<<

Swap2::Swap2(int nkv) {

  _nkv = nkv;

#ifdef USE_TARGET_OFFLOAD

  vkv = (double *)aligned_alloc(64, _nkv * sizeof(double));

  vec_vkzm = (double *)aligned_alloc(64, _nkv * _nkv * sizeof(double));

#pragma omp parallel for collapse(2)

  for (int i = 0; i < _nkv; i++) {

    for (int j = 0; j < _nkv; j++) {

      vkv[i] = 0.0;

      vec_vkzm[_nkv * i +j] = 0.0;

    }

  }

#else

  vkv = new double[_nkv]();

  vec_vkzm = new double[_nkv * _nkv]();

#endif // USE TARGET_OFFLOAD

  _last_vector = 0;

  _last_matrix = 0;

}



Swap2::~Swap2() {

#ifdef USE_TARGET_OFFLOAD

  free(vkv);

  free(vec_vkzm);

#else

  delete[] vkv;

  delete[] vec_vkzm;

#endif // USE_TARGET_OFFLOAD

}



Swap2* Swap2::from_binary(const std::string& file_name, const std::string& mode) {
@@ -578,19 +599,33 @@ TFRFME::TFRFME(int lmode, int lm, int nkv, int nxv, int nyv, int nzv) { 
  _exril = 0.0;



  // Array initialization

  xv = new double[nxv]();

  yv = new double[nyv]();

  zv = new double[nzv]();

  _nlmmt = _lm * (_lm + 2) * 2;

  _nrvc = _nxv * _nyv * _nzv;

  vec_wsum = new dcomplex[nrvc * nlmmt]();

#ifdef USE_TARGET_OFFLOAD

  xv = (double *)aligned_alloc(64, sizeof(double) * _nxv);

  yv = (double *)aligned_alloc(64, sizeof(double) * _nyv);

  zv = (double *)aligned_alloc(64, sizeof(double) * _nzv);

  vec_wsum = (dcomplex *)aligned_alloc(64, sizeof(dcomplex) * _nrvc * _nlmmt);

#else

  xv = new double[_nxv]();

  yv = new double[_nyv]();

  zv = new double[_nzv]();

  vec_wsum = new dcomplex[_nrvc * _nlmmt]();

#endif // USE_TARGET_OFFLOAD

}



TFRFME::~TFRFME() {

#ifdef USE_TARGET_OFFLOAD

  free(xv);

  free(yv);

  free(zv);

  free(vec_wsum);

#else

  delete[] xv;

  delete[] yv;

  delete[] zv;

  delete[] vec_wsum;

#endif

}



TFRFME* TFRFME::from_binary(const std::string& file_name, const std::string& mode) {