ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fixed bugs with hippo/gpu for single- and mixed- precisions

Browse Source
This commit is contained in:
Trung Nguyen
2023-01-24 23:55:30 -06:00
parent 6c63d7dcb9
commit b206b4d1f6
1 changed files with 86 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

61
src/GPU/pair_hippo_gpu.cpp
View File

@ -849,7 +849,8 @@ void PairHippoGPU::udirect2b(double **field, double **fieldp)
// field and fieldp may already have some nonzero values from kspace (udirect1) // field and fieldp may already have some nonzero values from kspace (udirect1)
int nlocal = atom->nlocal; int nlocal = atom->nlocal;
double *field_ptr = (double *)fieldp_pinned; if (tq_single) {
auto field_ptr = (float *)fieldp_pinned;
for (int i = 0; i < nlocal; i++) { for (int i = 0; i < nlocal; i++) {
int idx = 4*i; int idx = 4*i;
@ -858,7 +859,7 @@ void PairHippoGPU::udirect2b(double **field, double **fieldp)
field[i][2] += field_ptr[idx+2]; field[i][2] += field_ptr[idx+2];
} }
double* fieldp_ptr = (double *)fieldp_pinned; auto fieldp_ptr = (float *)fieldp_pinned;
fieldp_ptr += 4*inum; fieldp_ptr += 4*inum;
for (int i = 0; i < nlocal; i++) { for (int i = 0; i < nlocal; i++) {
int idx = 4*i; int idx = 4*i;
@ -867,6 +868,25 @@ void PairHippoGPU::udirect2b(double **field, double **fieldp)
fieldp[i][2] += fieldp_ptr[idx+2]; fieldp[i][2] += fieldp_ptr[idx+2];
} }
} else {
auto field_ptr = (double *)fieldp_pinned;
for (int i = 0; i < nlocal; i++) {
int idx = 4*i;
field[i][0] += field_ptr[idx];
field[i][1] += field_ptr[idx+1];
field[i][2] += field_ptr[idx+2];
}
auto fieldp_ptr = (double *)fieldp_pinned;
fieldp_ptr += 4*inum;
for (int i = 0; i < nlocal; i++) {
int idx = 4*i;
fieldp[i][0] += fieldp_ptr[idx];
fieldp[i][1] += fieldp_ptr[idx+1];
fieldp[i][2] += fieldp_ptr[idx+2];
}
}
} }
/* ---------------------------------------------------------------------- /* ----------------------------------------------------------------------
@ -1246,7 +1266,8 @@ void PairHippoGPU::fphi_uind(FFT_SCALAR ****grid, double **fdip_phi1,
&fdip_sum_phi_pinned); &fdip_sum_phi_pinned);
int nlocal = atom->nlocal; int nlocal = atom->nlocal;
double *_fdip_phi1_ptr = (double *)fdip_phi1_pinned; if (tq_single) {
auto _fdip_phi1_ptr = (float *)fdip_phi1_pinned;
for (int i = 0; i < nlocal; i++) { for (int i = 0; i < nlocal; i++) {
int n = i; int n = i;
for (int m = 0; m < 10; m++) { for (int m = 0; m < 10; m++) {
@ -1255,7 +1276,7 @@ void PairHippoGPU::fphi_uind(FFT_SCALAR ****grid, double **fdip_phi1,
} }
} }
double *_fdip_phi2_ptr = (double *)fdip_phi2_pinned; auto _fdip_phi2_ptr = (float *)fdip_phi2_pinned;
for (int i = 0; i < nlocal; i++) { for (int i = 0; i < nlocal; i++) {
int n = i; int n = i;
for (int m = 0; m < 10; m++) { for (int m = 0; m < 10; m++) {
@ -1264,7 +1285,7 @@ void PairHippoGPU::fphi_uind(FFT_SCALAR ****grid, double **fdip_phi1,
} }
} }
double *_fdip_sum_phi_ptr = (double *)fdip_sum_phi_pinned; auto _fdip_sum_phi_ptr = (float *)fdip_sum_phi_pinned;
for (int i = 0; i < nlocal; i++) { for (int i = 0; i < nlocal; i++) {
int n = i; int n = i;
for (int m = 0; m < 20; m++) { for (int m = 0; m < 20; m++) {
@ -1272,6 +1293,36 @@ void PairHippoGPU::fphi_uind(FFT_SCALAR ****grid, double **fdip_phi1,
n += nlocal; n += nlocal;
} }
} }
} else {
auto _fdip_phi1_ptr = (double *)fdip_phi1_pinned;
for (int i = 0; i < nlocal; i++) {
int n = i;
for (int m = 0; m < 10; m++) {
fdip_phi1[i][m] = _fdip_phi1_ptr[n];
n += nlocal;
}
}
auto _fdip_phi2_ptr = (double *)fdip_phi2_pinned;
for (int i = 0; i < nlocal; i++) {
int n = i;
for (int m = 0; m < 10; m++) {
fdip_phi2[i][m] = _fdip_phi2_ptr[n];
n += nlocal;
}
}
auto _fdip_sum_phi_ptr = (double *)fdip_sum_phi_pinned;
for (int i = 0; i < nlocal; i++) {
int n = i;
for (int m = 0; m < 20; m++) {
fdip_sum_phi[i][m] = _fdip_sum_phi_ptr[n];
n += nlocal;
}
}
}
} }
/* ---------------------------------------------------------------------- /* ----------------------------------------------------------------------