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

Write LAMMPS kernel with SIMD intrinsics and implement AVX512 with double-precision functions

Browse Source 
Signed-off-by: Rafael Ravedutti <rafaelravedutti@gmail.com>
This commit is contained in:
Rafael Ravedutti
2022-04-05 02:57:23 +02:00
parent af1756bfe4
commit ab2eb1ff50
6 changed files with 113 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

73
lammps/force_lj.c
View File

@@ -28,7 +28,10 @@
#include <atom.h>
#include <stats.h>
double computeForceLJFullNeigh(Parameter *param, Atom *atom, Neighbor *neighbor, Stats *stats) {
// TODO: Joint common files for gromacs and lammps variants
#include "../gromacs/includes/simd.h"
double computeForceLJFullNeigh_plain_c(Parameter *param, Atom *atom, Neighbor *neighbor, Stats *stats) {
int Nlocal = atom->Nlocal;
int* neighs;
#ifndef EXPLICIT_TYPES
@@ -184,3 +187,71 @@ double computeForceLJHalfNeigh(Parameter *param, Atom *atom, Neighbor *neighbor,
double E = getTimeStamp();
return E-S;
}
double computeForceLJFullNeigh_simd(Parameter *param, Atom *atom, Neighbor *neighbor, Stats *stats) {
int Nlocal = atom->Nlocal;
int* neighs;
MD_FLOAT cutforcesq = param->cutforce * param->cutforce;
MD_FLOAT sigma6 = param->sigma6;
MD_FLOAT epsilon = param->epsilon;
MD_SIMD_FLOAT cutforcesq_vec = simd_broadcast(cutforcesq);
MD_SIMD_FLOAT sigma6_vec = simd_broadcast(sigma6);
MD_SIMD_FLOAT eps_vec = simd_broadcast(epsilon);
MD_SIMD_FLOAT c48_vec = simd_broadcast(48.0);
MD_SIMD_FLOAT c05_vec = simd_broadcast(0.5);
for(int i = 0; i < Nlocal; i++) {
atom_fx(i) = 0.0;
atom_fy(i) = 0.0;
atom_fz(i) = 0.0;
}
double S = getTimeStamp();
LIKWID_MARKER_START("force");
#pragma omp parallel for
for(int i = 0; i < Nlocal; i++) {
neighs = &neighbor->neighbors[i * neighbor->maxneighs];
int numneighs = neighbor->numneigh[i];
MD_SIMD_INT numneighs_vec = simd_int_broadcast(numneighs);
MD_SIMD_FLOAT xtmp = simd_broadcast(atom_x(i));
MD_SIMD_FLOAT ytmp = simd_broadcast(atom_y(i));
MD_SIMD_FLOAT ztmp = simd_broadcast(atom_z(i));
MD_SIMD_FLOAT fix = simd_zero();
MD_SIMD_FLOAT fiy = simd_zero();
MD_SIMD_FLOAT fiz = simd_zero();
for(int k = 0; k < numneighs; k += VECTOR_WIDTH) {
// If the last iteration of this loop is separated from the rest, this mask can be set only there
MD_SIMD_MASK mask_numneighs = simd_mask_int_cond_lt(simd_int_add(simd_int_broadcast(k), simd_int_seq()), numneighs_vec);
MD_SIMD_INT j = simd_int_mask_load(&neighs[k], mask_numneighs);
#ifdef AOS
MD_SIMD_INT j3 = simd_int_add(simd_int_add(j, j), j); // j * 3
MD_SIMD_FLOAT delx = xtmp - simd_gather(j3, &(atom->x[0]), sizeof(MD_FLOAT));
MD_SIMD_FLOAT dely = ytmp - simd_gather(j3, &(atom->x[1]), sizeof(MD_FLOAT));
MD_SIMD_FLOAT delz = ztmp - simd_gather(j3, &(atom->x[2]), sizeof(MD_FLOAT));
#else
MD_SIMD_FLOAT delx = xtmp - simd_gather(j, atom->x, sizeof(MD_FLOAT));
MD_SIMD_FLOAT dely = ytmp - simd_gather(j, atom->y, sizeof(MD_FLOAT));
MD_SIMD_FLOAT delz = ztmp - simd_gather(j, atom->z, sizeof(MD_FLOAT));
#endif
MD_SIMD_FLOAT rsq = simd_fma(delx, delx, simd_fma(dely, dely, simd_mul(delz, delz)));
MD_SIMD_MASK cutoff_mask = simd_mask_and(mask_numneighs, simd_mask_cond_lt(rsq, cutforcesq_vec));
MD_SIMD_FLOAT sr2 = simd_reciprocal(rsq);
MD_SIMD_FLOAT sr6 = simd_mul(sr2, simd_mul(sr2, simd_mul(sr2, sigma6_vec)));
MD_SIMD_FLOAT force = simd_mul(c48_vec, simd_mul(sr6, simd_mul(simd_sub(sr6, c05_vec), simd_mul(sr2, eps_vec))));
fix = simd_masked_add(fix, simd_mul(delx, force), cutoff_mask);
fiy = simd_masked_add(fiy, simd_mul(dely, force), cutoff_mask);
fiz = simd_masked_add(fiz, simd_mul(delz, force), cutoff_mask);
}
atom_fx(i) += simd_h_reduce_sum(fix);
atom_fy(i) += simd_h_reduce_sum(fiy);
atom_fz(i) += simd_h_reduce_sum(fiz);
}
LIKWID_MARKER_STOP("force");
double E = getTimeStamp();
return E-S;
}

11
lammps/main.c
View File

@@ -45,10 +45,19 @@
#define HLINE "----------------------------------------------------------------------------\n"
extern double computeForceLJFullNeigh(Parameter*, Atom*, Neighbor*, Stats*);
extern double computeForceLJFullNeigh_plain_c(Parameter*, Atom*, Neighbor*, Stats*);
extern double computeForceLJFullNeigh_simd(Parameter*, Atom*, Neighbor*, Stats*);
extern double computeForceLJHalfNeigh(Parameter*, Atom*, Neighbor*, Stats*);
extern double computeForceEam(Eam*, Parameter*, Atom*, Neighbor*, Stats*);
#ifdef USE_SIMD_KERNEL
# define KERNEL_NAME "SIMD"
# define computeForceLJFullNeigh computeForceLJFullNeigh_simd
#else
# define KERNEL_NAME "plain-C"
# define computeForceLJFullNeigh computeForceLJFullNeigh_plain_c
#endif
double setup(Parameter *param, Eam *eam, Atom *atom, Neighbor *neighbor, Stats *stats) {
if(param->force_field == FF_EAM) { initEam(eam, param); }
double S, E;