moebiusband/MD-Bench
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix explicit types for CUDA and provide option to write initial state of system

Browse Source 
Signed-off-by: Rafael Ravedutti <rafaelravedutti@gmail.com>
This commit is contained in:
Rafael Ravedutti
2023-12-13 10:52:47 +01:00
parent da3b1dd53f
commit 02629612a9
9 changed files with 42 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
lammps/cuda/force.cu
View File

@@ -29,7 +29,7 @@ extern "C" {
}
// cuda kernel
__global__ void calc_force(DeviceAtom a, MD_FLOAT cutforcesq, MD_FLOAT sigma6, MD_FLOAT epsilon, int Nlocal, int neigh_maxneighs, int *neigh_neighbors, int *neigh_numneigh) {
__global__ void calc_force(DeviceAtom a, MD_FLOAT cutforcesq, MD_FLOAT sigma6, MD_FLOAT epsilon, int Nlocal, int neigh_maxneighs, int *neigh_neighbors, int *neigh_numneigh, int ntypes) {
const int i = blockIdx.x * blockDim.x + threadIdx.x;
if(i >= Nlocal) {
return;
@@ -46,6 +46,10 @@ __global__ void calc_force(DeviceAtom a, MD_FLOAT cutforcesq, MD_FLOAT sigma6, M
MD_FLOAT fiy = 0;
MD_FLOAT fiz = 0;
#ifdef EXPLICIT_TYPES
const int type_i = atom->type[i];
#endif
for(int k = 0; k < numneighs; k++) {
int j = neigh_neighbors[Nlocal * k + i];
MD_FLOAT delx = xtmp - atom_x(j);
@@ -55,7 +59,7 @@ __global__ void calc_force(DeviceAtom a, MD_FLOAT cutforcesq, MD_FLOAT sigma6, M
#ifdef EXPLICIT_TYPES
const int type_j = atom->type[j];
const int type_ij = type_i * atom->ntypes + type_j;
const int type_ij = type_i * ntypes + type_j;
const MD_FLOAT cutforcesq = atom->cutforcesq[type_ij];
const MD_FLOAT sigma6 = atom->sigma6[type_ij];
const MD_FLOAT epsilon = atom->epsilon[type_ij];
@@ -138,11 +142,9 @@ void initialIntegrate_cuda(bool reneigh, Parameter *param, Atom *atom) {
double computeForceLJFullNeigh_cuda(Parameter *param, Atom *atom, Neighbor *neighbor) {
const int num_threads_per_block = get_cuda_num_threads();
int Nlocal = atom->Nlocal;
#ifndef EXPLICIT_TYPES
MD_FLOAT cutforcesq = param->cutforce * param->cutforce;
MD_FLOAT sigma6 = param->sigma6;
MD_FLOAT epsilon = param->epsilon;
#endif
/*
int nDevices;
@@ -165,7 +167,7 @@ double computeForceLJFullNeigh_cuda(Parameter *param, Atom *atom, Neighbor *neig
double S = getTimeStamp();
LIKWID_MARKER_START("force");
calc_force <<< num_blocks, num_threads_per_block >>> (atom->d_atom, cutforcesq, sigma6, epsilon, Nlocal, neighbor->maxneighs, neighbor->d_neighbor.neighbors, neighbor->d_neighbor.numneigh);
calc_force <<< num_blocks, num_threads_per_block >>> (atom->d_atom, cutforcesq, sigma6, epsilon, Nlocal, neighbor->maxneighs, neighbor->d_neighbor.neighbors, neighbor->d_neighbor.numneigh, atom->ntypes);
cuda_assert("calc_force", cudaPeekAtLastError());
cuda_assert("calc_force", cudaDeviceSynchronize());
cudaProfilerStop();

6
lammps/cuda/neighbor.cu
View File

@@ -120,7 +120,7 @@ __global__ void binatoms_kernel(DeviceAtom a, int nall, int* bincount, int* bins
__global__ void compute_neighborhood(
DeviceAtom a, DeviceNeighbor neigh, Neighbor_params np, int nlocal, int maxneighs, int nstencil, int* stencil,
int* bins, int atoms_per_bin, int *bincount, int *new_maxneighs, MD_FLOAT cutneighsq) {
int* bins, int atoms_per_bin, int *bincount, int *new_maxneighs, MD_FLOAT cutneighsq, int ntypes) {
const int i = blockIdx.x * blockDim.x + threadIdx.x;
if(i >= nlocal) {
@@ -157,7 +157,7 @@ __global__ void compute_neighborhood(
#ifdef EXPLICIT_TYPES
int type_j = atom->type[j];
const MD_FLOAT cutoff = atom->cutneighsq[type_i * atom->ntypes + type_j];
const MD_FLOAT cutoff = atom->cutneighsq[type_i * ntypes + type_j];
#else
const MD_FLOAT cutoff = cutneighsq;
#endif
@@ -269,7 +269,7 @@ void buildNeighbor_cuda(Atom *atom, Neighbor *neighbor) {
np, atom->Nlocal, neighbor->maxneighs, nstencil, c_stencil,
c_binning.bins, c_binning.atoms_per_bin, c_binning.bincount,
c_new_maxneighs,
cutneighsq);
cutneighsq, atom->ntypes);
cuda_assert("compute_neighborhood", cudaPeekAtLastError());
cuda_assert("compute_neighborhood", cudaDeviceSynchronize());