Rough rewrite to execute outer loop of force calculation in parallel, not inner loop

This commit is contained in:
Maximilian Gaul 2021-11-14 10:02:23 +01:00
parent e2fd1a0476
commit 2e5d973f7d

View File

@ -49,42 +49,59 @@ void checkError(const char *msg, cudaError_t err)
// cuda kernel // cuda kernel
__global__ void calc_force( __global__ void calc_force(
Atom a, Atom a,
MD_FLOAT xtmp, MD_FLOAT ytmp, MD_FLOAT ztmp,
MD_FLOAT *fix, MD_FLOAT *fiy, MD_FLOAT *fiz,
MD_FLOAT cutforcesq, MD_FLOAT sigma6, MD_FLOAT epsilon, MD_FLOAT cutforcesq, MD_FLOAT sigma6, MD_FLOAT epsilon,
int i, int numneighs, int *neighs) { int Nlocal, int neigh_maxneighs, int *neigh_neighbors, int *neigh_numneigh) {
// Calculate idx k from thread information const int i = blockIdx.x * blockDim.x + threadIdx.x;
const long long k = blockIdx.x * blockDim.x + threadIdx.x; if( i >= Nlocal ) {
if( k >= numneighs ) {
return; return;
} }
Atom *atom = &a; Atom *atom = &a;
const int j = neighs[k]; int *neighs = &neigh_neighbors[i * neigh_maxneighs];
MD_FLOAT delx = xtmp - atom_x(j); int numneighs = neigh_numneigh[i];
MD_FLOAT dely = ytmp - atom_y(j);
MD_FLOAT delz = ztmp - atom_z(j); MD_FLOAT xtmp = atom_x(i);
MD_FLOAT rsq = delx * delx + dely * dely + delz * delz; MD_FLOAT ytmp = atom_y(i);
MD_FLOAT ztmp = atom_z(i);
MD_FLOAT *fx = atom->fx;
MD_FLOAT *fy = atom->fy;
MD_FLOAT *fz = atom->fz;
MD_FLOAT fix = 0;
MD_FLOAT fiy = 0;
MD_FLOAT fiz = 0;
for(int k = 0; k < numneighs; k++) {
int j = neighs[k];
MD_FLOAT delx = xtmp - atom_x(j);
MD_FLOAT dely = ytmp - atom_y(j);
MD_FLOAT delz = ztmp - atom_z(j);
MD_FLOAT rsq = delx * delx + dely * dely + delz * delz;
#ifdef EXPLICIT_TYPES #ifdef EXPLICIT_TYPES
const int type_i = atom->type[i]; const int type_j = atom->type[j];
const int type_j = atom->type[j]; const int type_ij = type_i * atom->ntypes + type_j;
const int type_ij = type_i * atom->ntypes + type_j; const MD_FLOAT cutforcesq = atom->cutforcesq[type_ij];
const MD_FLOAT cutforcesq = atom->cutforcesq[type_ij]; const MD_FLOAT sigma6 = atom->sigma6[type_ij];
const MD_FLOAT sigma6 = atom->sigma6[type_ij]; const MD_FLOAT epsilon = atom->epsilon[type_ij];
const MD_FLOAT epsilon = atom->epsilon[type_ij];
#endif #endif
if(rsq < cutforcesq) { if(rsq < cutforcesq) {
MD_FLOAT sr2 = 1.0 / rsq; MD_FLOAT sr2 = 1.0 / rsq;
MD_FLOAT sr6 = sr2 * sr2 * sr2 * sigma6; MD_FLOAT sr6 = sr2 * sr2 * sr2 * sigma6;
MD_FLOAT force = 48.0 * sr6 * (sr6 - 0.5) * sr2 * epsilon; MD_FLOAT force = 48.0 * sr6 * (sr6 - 0.5) * sr2 * epsilon;
fix[k] += delx * force; fix += delx * force;
fiy[k] += dely * force; fiy += dely * force;
fiz[k] += delz * force; fiz += delz * force;
}
} }
fx[i] += fix;
fy[i] += fiy;
fz[i] += fiz;
} }
extern "C" { extern "C" {
@ -96,7 +113,6 @@ double computeForce(
) )
{ {
int Nlocal = atom->Nlocal; int Nlocal = atom->Nlocal;
int* neighs;
MD_FLOAT* fx = atom->fx; MD_FLOAT* fx = atom->fx;
MD_FLOAT* fy = atom->fy; MD_FLOAT* fy = atom->fy;
MD_FLOAT* fz = atom->fz; MD_FLOAT* fz = atom->fz;
@ -122,88 +138,63 @@ double computeForce(
// HINT: Run with cuda-memcheck ./MDBench-NVCC in case of error // HINT: Run with cuda-memcheck ./MDBench-NVCC in case of error
// HINT: Only works for data layout = AOS!!! // HINT: Only works for data layout = AOS!!!
checkError( "Malloc1", cudaMalloc((void**)&(c_atom.x), sizeof(MD_FLOAT) * atom->Nmax * 3) ); checkError( "c_atom.x malloc", cudaMalloc((void**)&(c_atom.x), sizeof(MD_FLOAT) * atom->Nmax * 3) );
checkError( "Memcpy1", cudaMemcpy(c_atom.x, atom->x, sizeof(MD_FLOAT) * atom->Nmax * 3, cudaMemcpyHostToDevice) ); checkError( "c_atom.x memcpy", cudaMemcpy(c_atom.x, atom->x, sizeof(MD_FLOAT) * atom->Nmax * 3, cudaMemcpyHostToDevice) );
checkError( "Malloc2", cudaMalloc((void**)&(c_atom.type), sizeof(int) * atom->Nmax) ); checkError( "c_atom.fx malloc", cudaMalloc((void**)&(c_atom.fx), sizeof(MD_FLOAT) * Nlocal) );
checkError( "Memcpy2", cudaMemcpy(c_atom.type, atom->type, sizeof(int) * atom->Nmax, cudaMemcpyHostToDevice) ); checkError( "c_atom.fx memcpy", cudaMemcpy(c_atom.fx, fx, sizeof(MD_FLOAT) * Nlocal, cudaMemcpyHostToDevice) );
checkError( "Malloc3", cudaMalloc((void**)&(c_atom.epsilon), sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes) ); checkError( "c_atom.fy malloc", cudaMalloc((void**)&(c_atom.fy), sizeof(MD_FLOAT) * Nlocal) );
checkError( "Memcpy3", cudaMemcpy(c_atom.epsilon, atom->epsilon, sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes, cudaMemcpyHostToDevice) ); checkError( "c_atom.fy memcpy", cudaMemcpy(c_atom.fy, fy, sizeof(MD_FLOAT) * Nlocal, cudaMemcpyHostToDevice) );
checkError( "Malloc4", cudaMalloc((void**)&(c_atom.sigma6), sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes) ); checkError( "c_atom.fz malloc", cudaMalloc((void**)&(c_atom.fz), sizeof(MD_FLOAT) * Nlocal) );
checkError( "Memcpy4", cudaMemcpy(c_atom.sigma6, atom->sigma6, sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes, cudaMemcpyHostToDevice) ); checkError( "c_atom.fz memcpy", cudaMemcpy(c_atom.fz, fz, sizeof(MD_FLOAT) * Nlocal, cudaMemcpyHostToDevice) );
checkError( "Malloc5", cudaMalloc((void**)&(c_atom.cutforcesq), sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes) ); checkError( "c_atom.type malloc", cudaMalloc((void**)&(c_atom.type), sizeof(int) * atom->Nmax) );
checkError( "Memcpy5", cudaMemcpy(c_atom.cutforcesq, atom->cutforcesq, sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes, cudaMemcpyHostToDevice) ); checkError( "c_atom.type memcpy", cudaMemcpy(c_atom.type, atom->type, sizeof(int) * atom->Nmax, cudaMemcpyHostToDevice) );
checkError( "c_atom.epsilon malloc", cudaMalloc((void**)&(c_atom.epsilon), sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes) );
checkError( "c_atom.epsilon memcpy", cudaMemcpy(c_atom.epsilon, atom->epsilon, sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes, cudaMemcpyHostToDevice) );
checkError( "c_atom.sigma6 malloc", cudaMalloc((void**)&(c_atom.sigma6), sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes) );
checkError( "c_atom.sigma6 memcpy", cudaMemcpy(c_atom.sigma6, atom->sigma6, sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes, cudaMemcpyHostToDevice) );
checkError( "c_atom.cutforcesq malloc", cudaMalloc((void**)&(c_atom.cutforcesq), sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes) );
checkError( "c_atom.cutforcesq memcpy", cudaMemcpy(c_atom.cutforcesq, atom->cutforcesq, sizeof(MD_FLOAT) * atom->ntypes * atom->ntypes, cudaMemcpyHostToDevice) );
int *c_neighs;
checkError( "c_neighs malloc", cudaMalloc((void**)&c_neighs, sizeof(int) * Nlocal * neighbor->maxneighs) );
checkError( "c_neighs memcpy", cudaMemcpy(c_neighs, neighbor->neighbors, sizeof(int) * Nlocal * neighbor->maxneighs, cudaMemcpyHostToDevice) );
int *c_neigh_numneigh;
checkError( "c_neigh_numneigh malloc", cudaMalloc((void**)&c_neigh_numneigh, sizeof(int) * Nlocal) );
checkError( "c_neigh_numneigh memcpy", cudaMemcpy(c_neigh_numneigh, neighbor->numneigh, sizeof(int) * Nlocal, cudaMemcpyHostToDevice) );
const int num_blocks = 1024;
const int num_threads_per_block = ceil((float)Nlocal / (float)num_blocks);
double S = getTimeStamp(); double S = getTimeStamp();
LIKWID_MARKER_START("force"); LIKWID_MARKER_START("force");
#pragma omp parallel for calc_force <<< num_blocks, num_threads_per_block >>> (c_atom, cutforcesq, sigma6, epsilon, Nlocal, neighbor->maxneighs, c_neighs, c_neigh_numneigh);
for(int i = 0; i < Nlocal; i++) {
neighs = &neighbor->neighbors[i * neighbor->maxneighs];
int numneighs = neighbor->numneigh[i];
MD_FLOAT xtmp = atom_x(i);
MD_FLOAT ytmp = atom_y(i);
MD_FLOAT ztmp = atom_z(i);
#ifdef EXPLICIT_TYPES checkError( "PeekAtLastError", cudaPeekAtLastError() );
const int type_i = atom->type[i]; checkError( "DeviceSync", cudaDeviceSynchronize() );
#endif
int *c_neighs; // copy results in c_atom.fx/fy/fz to atom->fx/fy/fz
checkError( "c_neighs malloc", cudaMalloc((void**)&c_neighs, sizeof(int) * numneighs) ); cudaMemcpy(atom->fx, c_atom.fx, sizeof(MD_FLOAT) * Nlocal, cudaMemcpyDeviceToHost);
checkError( "c_neighs memcpy", cudaMemcpy(c_neighs, neighs, sizeof(int) * numneighs, cudaMemcpyHostToDevice) ); cudaMemcpy(atom->fy, c_atom.fy, sizeof(MD_FLOAT) * Nlocal, cudaMemcpyDeviceToHost);
cudaMemcpy(atom->fz, c_atom.fz, sizeof(MD_FLOAT) * Nlocal, cudaMemcpyDeviceToHost);
const int num_blocks = 64;
const int num_threads_per_block = ceil((float)numneighs / (float)num_blocks);
// printf("numneighs: %d => num-blocks: %d, num_threads_per_block => %d\r\n", numneighs, num_blocks, num_threads_per_block);
MD_FLOAT *c_fix, *c_fiy, *c_fiz;
checkError( "c_fix malloc", cudaMalloc((void**)&c_fix, sizeof(MD_FLOAT) * numneighs) );
checkError( "c_fiy malloc", cudaMalloc((void**)&c_fiy, sizeof(MD_FLOAT) * numneighs) );
checkError( "c_fiz malloc", cudaMalloc((void**)&c_fiz, sizeof(MD_FLOAT) * numneighs) );
checkError( "c_fix memset", cudaMemset(c_fix, 0, sizeof(MD_FLOAT) * numneighs) );
checkError( "c_fiy memset", cudaMemset(c_fiy, 0, sizeof(MD_FLOAT) * numneighs) );
checkError( "c_fiz memset", cudaMemset(c_fiz, 0, sizeof(MD_FLOAT) * numneighs) );
// launch cuda kernel
calc_force <<< num_blocks, num_threads_per_block >>> (c_atom, xtmp, ytmp, ztmp, c_fix, c_fiy, c_fiz, cutforcesq, sigma6, epsilon, i, numneighs, c_neighs);
checkError( "PeekAtLastError", cudaPeekAtLastError() );
checkError( "DeviceSync", cudaDeviceSynchronize() );
MD_FLOAT *d_fix = (MD_FLOAT*)malloc(sizeof(MD_FLOAT) * numneighs);
MD_FLOAT *d_fiy = (MD_FLOAT*)malloc(sizeof(MD_FLOAT) * numneighs);
MD_FLOAT *d_fiz = (MD_FLOAT*)malloc(sizeof(MD_FLOAT) * numneighs);
// sum result
checkError( "d_fix copy to host", cudaMemcpy(d_fix, c_fix, sizeof(MD_FLOAT) * numneighs, cudaMemcpyDeviceToHost) );
checkError( "d_fiy copy to host", cudaMemcpy(d_fiy, c_fiy, sizeof(MD_FLOAT) * numneighs, cudaMemcpyDeviceToHost) );
checkError( "d_fiz copy to host", cudaMemcpy(d_fiz, c_fiz, sizeof(MD_FLOAT) * numneighs, cudaMemcpyDeviceToHost) );
for(int k = 0; k < numneighs; k++) {
fx[i] += d_fix[k];
fy[i] += d_fiy[k];
fz[i] += d_fiz[k];
}
checkError( "cudaFree c_fix", cudaFree(c_fix) );
checkError( "cudaFree c_fiy", cudaFree(c_fiy) );
checkError( "cudaFree c_fiz", cudaFree(c_fiz) );
checkError( "cudaFree c_neighs", cudaFree(c_neighs) );
free(d_fix);
free(d_fiy);
free(d_fiz);
}
cudaFree(c_atom.x); cudaFree(c_atom.x);
cudaFree(c_atom.fx); cudaFree(c_atom.fy); cudaFree(c_atom.fz);
cudaFree(c_atom.type); cudaFree(c_atom.type);
cudaFree(c_atom.epsilon); cudaFree(c_atom.epsilon);
cudaFree(c_atom.sigma6); cudaFree(c_atom.sigma6);
cudaFree(c_atom.cutforcesq); cudaFree(c_atom.cutforcesq);
cudaFree(c_neighs); cudaFree(c_neigh_numneigh);
LIKWID_MARKER_STOP("force"); LIKWID_MARKER_STOP("force");
double E = getTimeStamp(); double E = getTimeStamp();