#include <math.h>
#include <string.h>
+#include <assert.h>
+
#include "sysstuff.h"
-#include "smalloc.h"
+#include "gromacs/utility/smalloc.h"
+#include "types/commrec.h"
#include "macros.h"
#include "gromacs/math/utilities.h"
#include "vec.h"
#include "nbnxn_search.h"
#include "gmx_omp_nthreads.h"
#include "nrnb.h"
+#include "ns.h"
#include "gromacs/fileio/gmxfio.h"
}
}
+/* Initializes a single nbnxn_pairlist_t data structure */
+static void nbnxn_init_pairlist_fep(t_nblist *nl)
+{
+ nl->type = GMX_NBLIST_INTERACTION_FREE_ENERGY;
+ nl->igeometry = GMX_NBLIST_GEOMETRY_PARTICLE_PARTICLE;
+ /* The interaction functions are set in the free energy kernel fuction */
+ nl->ivdw = -1;
+ nl->ivdwmod = -1;
+ nl->ielec = -1;
+ nl->ielecmod = -1;
+
+ nl->maxnri = 0;
+ nl->maxnrj = 0;
+ nl->nri = 0;
+ nl->nrj = 0;
+ nl->iinr = NULL;
+ nl->gid = NULL;
+ nl->shift = NULL;
+ nl->jindex = NULL;
+ nl->jjnr = NULL;
+ nl->excl_fep = NULL;
+
+}
+
void nbnxn_init_search(nbnxn_search_t * nbs_ptr,
ivec *n_dd_cells,
gmx_domdec_zones_t *zones,
+ gmx_bool bFEP,
int nthread_max)
{
nbnxn_search_t nbs;
snew(nbs, 1);
*nbs_ptr = nbs;
+ nbs->bFEP = bFEP;
+
nbs->DomDec = (n_dd_cells != NULL);
clear_ivec(nbs->dd_dim);
nbs->work[t].cxy_na_nalloc = 0;
nbs->work[t].sort_work = NULL;
nbs->work[t].sort_work_nalloc = 0;
+
+ snew(nbs->work[t].nbl_fep, 1);
+ nbnxn_init_pairlist_fep(nbs->work[t].nbl_fep);
}
/* Initialize detailed nbsearch cycle counting */
{
rvec size;
+ if (n == 0)
+ {
+ /* To avoid zero density we use a minimum of 1 atom */
+ n = 1;
+ }
+
rvec_sub(corner1, corner0, size);
return n/(size[XX]*size[YY]*size[ZZ]);
if (n > grid->na_sc)
{
+ assert(atom_density > 0);
+
/* target cell length */
if (grid->bSimple)
{
}
srenew(grid->flags, grid->nc_nalloc);
+ if (nbs->bFEP)
+ {
+ srenew(grid->fep, grid->nc_nalloc*grid->na_sc/grid->na_c);
+ }
}
copy_rvec(corner0, grid->c0);
int subc, s, a, n1, n2, a_lj_max, i, j;
int sort1[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
int sort2[NBNXN_NA_SC_MAX/GPU_NSUBCELL];
- gmx_bool haveQ;
+ gmx_bool haveQ, bFEP;
*flags = 0;
}
}
- /* If we don't have atom with LJ, there's nothing to sort */
+ /* If we don't have atoms with LJ, there's nothing to sort */
if (n1 > 0)
{
*flags |= NBNXN_CI_DO_LJ(subc);
grid->flags+(a0>>grid->na_c_2log)-grid->cell0);
}
+ if (nbs->bFEP)
+ {
+ /* Set the fep flag for perturbed atoms in this (sub-)cell */
+ int c, at;
+
+ /* The grid-local cluster/(sub-)cell index */
+ c = (a0 >> grid->na_c_2log) - grid->cell0*(grid->bSimple ? 1 : GPU_NSUBCELL);
+ grid->fep[c] = 0;
+ for (at = a0; at < a1; at++)
+ {
+ if (nbs->a[at] >= 0 && GET_CGINFO_FEP(atinfo[nbs->a[at]]))
+ {
+ grid->fep[c] |= (1 << (at - a0));
+ }
+ }
+ }
+
/* Now we have sorted the atoms, set the cell indices */
for (a = a0; a < a1; a++)
{
}
/* Sort the super-cell columns along z into the sub-cells. */
-#pragma omp parallel for num_threads(nbs->nthread_max) schedule(static)
- for (thread = 0; thread < nbs->nthread_max; thread++)
+#pragma omp parallel for num_threads(nthread) schedule(static)
+ for (thread = 0; thread < nthread; thread++)
{
if (grid->bSimple)
{
nbs->ePBC = ePBC;
copy_mat(box, nbs->box);
- if (atom_density >= 0)
+ /* Avoid zero density */
+ if (atom_density > 0)
{
grid->atom_density = atom_density;
}
* for the local atoms (dd_zone=0).
*/
nbs->natoms_nonlocal = a1 - nmoved;
+
+ if (debug)
+ {
+ fprintf(debug, "natoms_local = %5d atom_density = %5.1f\n",
+ nbs->natoms_local, grid->atom_density);
+ }
}
else
{
nbs->natoms_nonlocal = max(nbs->natoms_nonlocal, a1);
}
+ /* We always use the home zone (grid[0]) for setting the cell size,
+ * since determining densities for non-local zones is difficult.
+ */
nc_max_grid = set_grid_size_xy(nbs, grid,
dd_zone, n-nmoved, corner0, corner1,
nbs->grid[0].atom_density);
float *bbcz;
nbnxn_bb_t *bb;
int ncd, sc;
+ int nthreads gmx_unused;
grid = &nbs->grid[0];
bbcz = grid->bbcz_simple;
bb = grid->bb_simple;
-#pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
+ nthreads = gmx_omp_nthreads_get(emntPairsearch);
+#pragma omp parallel for num_threads(nthreads) schedule(static)
for (sc = 0; sc < grid->nc; sc++)
{
int c, tx, na;
}
/* Returns the i-interaction mask of the j sub-cell for index cj_ind */
-static unsigned nbl_imask0(const nbnxn_pairlist_t *nbl, int cj_ind)
+static unsigned int nbl_imask0(const nbnxn_pairlist_t *nbl, int cj_ind)
{
return nbl->cj4[cj_ind >> NBNXN_GPU_JGROUP_SIZE_2LOG].imei[0].imask;
}
}
snew(nbl_list->nbl, nbl_list->nnbl);
+ snew(nbl_list->nbl_fep, nbl_list->nnbl);
/* Execute in order to avoid memory interleaving between threads */
#pragma omp parallel for num_threads(nbl_list->nnbl) schedule(static)
for (i = 0; i < nbl_list->nnbl; i++)
{
nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, NULL, NULL);
}
+
+ snew(nbl_list->nbl_fep[i], 1);
+ nbnxn_init_pairlist_fep(nbl_list->nbl_fep[i]);
}
}
}
/* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp,
- * allocates extra memory, if necessary.
+ * generates a new element and allocates extra memory, if necessary.
*/
static void get_nbl_exclusions_1(nbnxn_pairlist_t *nbl, int cj4,
int warp, nbnxn_excl_t **excl)
}
/* Returns pointers to the exclusion mask for cj4-unit cj4 for both warps,
- * allocates extra memory, if necessary.
+ * generates a new element and allocates extra memory, if necessary.
*/
static void get_nbl_exclusions_2(nbnxn_pairlist_t *nbl, int cj4,
nbnxn_excl_t **excl_w0,
int npair;
int cjo, ci1, ci, cj, cj_gl;
int cj4_ind, cj_offset;
- unsigned imask;
+ unsigned int imask;
nbnxn_cj4_t *cj4;
#ifdef NBNXN_BBXXXX
const float *pbb_ci;
}
}
+/* Add a new i-entry to the FEP list and copy the i-properties */
+static gmx_inline void fep_list_new_nri_copy(t_nblist *nlist)
+{
+ /* Add a new i-entry */
+ nlist->nri++;
+
+ assert(nlist->nri < nlist->maxnri);
+
+ /* Duplicate the last i-entry, except for jindex, which continues */
+ nlist->iinr[nlist->nri] = nlist->iinr[nlist->nri-1];
+ nlist->shift[nlist->nri] = nlist->shift[nlist->nri-1];
+ nlist->gid[nlist->nri] = nlist->gid[nlist->nri-1];
+ nlist->jindex[nlist->nri] = nlist->nrj;
+}
+
+/* For load balancing of the free-energy lists over threads, we set
+ * the maximum nrj size of an i-entry to 40. This leads to good
+ * load balancing in the worst case scenario of a single perturbed
+ * particle on 16 threads, while not introducing significant overhead.
+ * Note that half of the perturbed pairs will anyhow end up in very small lists,
+ * since non perturbed i-particles will see few perturbed j-particles).
+ */
+const int max_nrj_fep = 40;
+
+/* Exclude the perturbed pairs from the Verlet list. This is only done to avoid
+ * singularities for overlapping particles (0/0), since the charges and
+ * LJ parameters have been zeroed in the nbnxn data structure.
+ * Simultaneously make a group pair list for the perturbed pairs.
+ */
+static void make_fep_list(const nbnxn_search_t nbs,
+ const nbnxn_atomdata_t *nbat,
+ nbnxn_pairlist_t *nbl,
+ gmx_bool bDiagRemoved,
+ nbnxn_ci_t *nbl_ci,
+ const nbnxn_grid_t *gridi,
+ const nbnxn_grid_t *gridj,
+ t_nblist *nlist)
+{
+ int ci, cj_ind_start, cj_ind_end, cj_ind, cja, cjr;
+ int nri_max;
+ int ngid, gid_i = 0, gid_j, gid;
+ int egp_shift, egp_mask;
+ int gid_cj = 0;
+ int i, j, ind_i, ind_j, ai, aj;
+ int nri;
+ gmx_bool bFEP_i, bFEP_i_all;
+
+ if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
+ {
+ /* Empty list */
+ return;
+ }
+
+ ci = nbl_ci->ci;
+
+ cj_ind_start = nbl_ci->cj_ind_start;
+ cj_ind_end = nbl_ci->cj_ind_end;
+
+ /* In worst case we have alternating energy groups
+ * and create #atom-pair lists, which means we need the size
+ * of a cluster pair (na_ci*na_cj) times the number of cj's.
+ */
+ nri_max = nbl->na_ci*nbl->na_cj*(cj_ind_end - cj_ind_start);
+ if (nlist->nri + nri_max > nlist->maxnri)
+ {
+ nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
+ reallocate_nblist(nlist);
+ }
+
+ ngid = nbat->nenergrp;
+
+ if (ngid*gridj->na_cj > sizeof(gid_cj)*8)
+ {
+ gmx_fatal(FARGS, "The Verlet scheme with %dx%d kernels and free-energy only supports up to %d energy groups",
+ gridi->na_c, gridj->na_cj, (sizeof(gid_cj)*8)/gridj->na_cj);
+ }
+
+ egp_shift = nbat->neg_2log;
+ egp_mask = (1<<nbat->neg_2log) - 1;
+
+ /* Loop over the atoms in the i sub-cell */
+ bFEP_i_all = TRUE;
+ for (i = 0; i < nbl->na_ci; i++)
+ {
+ ind_i = ci*nbl->na_ci + i;
+ ai = nbs->a[ind_i];
+ if (ai >= 0)
+ {
+ nri = nlist->nri;
+ nlist->jindex[nri+1] = nlist->jindex[nri];
+ nlist->iinr[nri] = ai;
+ /* The actual energy group pair index is set later */
+ nlist->gid[nri] = 0;
+ nlist->shift[nri] = nbl_ci->shift & NBNXN_CI_SHIFT;
+
+ bFEP_i = gridi->fep[ci - gridi->cell0] & (1 << i);
+
+ bFEP_i_all = bFEP_i_all && bFEP_i;
+
+ if ((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj > nlist->maxnrj)
+ {
+ nlist->maxnrj = over_alloc_small((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj);
+ srenew(nlist->jjnr, nlist->maxnrj);
+ srenew(nlist->excl_fep, nlist->maxnrj);
+ }
+
+ if (ngid > 1)
+ {
+ gid_i = (nbat->energrp[ci] >> (egp_shift*i)) & egp_mask;
+ }
+
+ for (cj_ind = cj_ind_start; cj_ind < cj_ind_end; cj_ind++)
+ {
+ unsigned int fep_cj;
+
+ cja = nbl->cj[cj_ind].cj;
+
+ if (gridj->na_cj == gridj->na_c)
+ {
+ cjr = cja - gridj->cell0;
+ fep_cj = gridj->fep[cjr];
+ if (ngid > 1)
+ {
+ gid_cj = nbat->energrp[cja];
+ }
+ }
+ else if (2*gridj->na_cj == gridj->na_c)
+ {
+ cjr = cja - gridj->cell0*2;
+ /* Extract half of the ci fep/energrp mask */
+ fep_cj = (gridj->fep[cjr>>1] >> ((cjr&1)*gridj->na_cj)) & ((1<<gridj->na_cj) - 1);
+ if (ngid > 1)
+ {
+ gid_cj = nbat->energrp[cja>>1] >> ((cja&1)*gridj->na_cj*egp_shift) & ((1<<(gridj->na_cj*egp_shift)) - 1);
+ }
+ }
+ else
+ {
+ cjr = cja - (gridj->cell0>>1);
+ /* Combine two ci fep masks/energrp */
+ fep_cj = gridj->fep[cjr*2] + (gridj->fep[cjr*2+1] << gridj->na_c);
+ if (ngid > 1)
+ {
+ gid_cj = nbat->energrp[cja*2] + (nbat->energrp[cja*2+1] << (gridj->na_c*egp_shift));
+ }
+ }
+
+ if (bFEP_i || fep_cj != 0)
+ {
+ for (j = 0; j < nbl->na_cj; j++)
+ {
+ /* Is this interaction perturbed and not excluded? */
+ ind_j = cja*nbl->na_cj + j;
+ aj = nbs->a[ind_j];
+ if (aj >= 0 &&
+ (bFEP_i || (fep_cj & (1 << j))) &&
+ (!bDiagRemoved || ind_j >= ind_i))
+ {
+ if (ngid > 1)
+ {
+ gid_j = (gid_cj >> (j*egp_shift)) & egp_mask;
+ gid = GID(gid_i, gid_j, ngid);
+
+ if (nlist->nrj > nlist->jindex[nri] &&
+ nlist->gid[nri] != gid)
+ {
+ /* Energy group pair changed: new list */
+ fep_list_new_nri_copy(nlist);
+ nri = nlist->nri;
+ }
+ nlist->gid[nri] = gid;
+ }
+
+ if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
+ {
+ fep_list_new_nri_copy(nlist);
+ nri = nlist->nri;
+ }
+
+ /* Add it to the FEP list */
+ nlist->jjnr[nlist->nrj] = aj;
+ nlist->excl_fep[nlist->nrj] = (nbl->cj[cj_ind].excl >> (i*nbl->na_cj + j)) & 1;
+ nlist->nrj++;
+
+ /* Exclude it from the normal list.
+ * Note that the charge has been set to zero,
+ * but we need to avoid 0/0, as perturbed atoms
+ * can be on top of each other.
+ */
+ nbl->cj[cj_ind].excl &= ~(1U << (i*nbl->na_cj + j));
+ }
+ }
+ }
+ }
+
+ if (nlist->nrj > nlist->jindex[nri])
+ {
+ /* Actually add this new, non-empty, list */
+ nlist->nri++;
+ nlist->jindex[nlist->nri] = nlist->nrj;
+ }
+ }
+ }
+
+ if (bFEP_i_all)
+ {
+ /* All interactions are perturbed, we can skip this entry */
+ nbl_ci->cj_ind_end = cj_ind_start;
+ }
+}
+
+/* Return the index of atom a within a cluster */
+static gmx_inline int cj_mod_cj4(int cj)
+{
+ return cj & (NBNXN_GPU_JGROUP_SIZE - 1);
+}
+
+/* Convert a j-cluster to a cj4 group */
+static gmx_inline int cj_to_cj4(int cj)
+{
+ return cj >> NBNXN_GPU_JGROUP_SIZE_2LOG;
+}
+
+/* Return the index of an j-atom within a warp */
+static gmx_inline int a_mod_wj(int a)
+{
+ return a & (NBNXN_GPU_CLUSTER_SIZE/2 - 1);
+}
+
+/* As make_fep_list above, but for super/sub lists. */
+static void make_fep_list_supersub(const nbnxn_search_t nbs,
+ const nbnxn_atomdata_t *nbat,
+ nbnxn_pairlist_t *nbl,
+ gmx_bool bDiagRemoved,
+ const nbnxn_sci_t *nbl_sci,
+ real shx,
+ real shy,
+ real shz,
+ real rlist_fep2,
+ const nbnxn_grid_t *gridi,
+ const nbnxn_grid_t *gridj,
+ t_nblist *nlist)
+{
+ int sci, cj4_ind_start, cj4_ind_end, cj4_ind, gcj, cjr;
+ int nri_max;
+ int c, c_abs;
+ int i, j, ind_i, ind_j, ai, aj;
+ int nri;
+ gmx_bool bFEP_i;
+ real xi, yi, zi;
+ const nbnxn_cj4_t *cj4;
+
+ if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
+ {
+ /* Empty list */
+ return;
+ }
+
+ sci = nbl_sci->sci;
+
+ cj4_ind_start = nbl_sci->cj4_ind_start;
+ cj4_ind_end = nbl_sci->cj4_ind_end;
+
+ /* Here we process one super-cell, max #atoms na_sc, versus a list
+ * cj4 entries, each with max NBNXN_GPU_JGROUP_SIZE cj's, each
+ * of size na_cj atoms.
+ * On the GPU we don't support energy groups (yet).
+ * So for each of the na_sc i-atoms, we need max one FEP list
+ * for each max_nrj_fep j-atoms.
+ */
+ nri_max = nbl->na_sc*nbl->na_cj*(1 + ((cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE)/max_nrj_fep);
+ if (nlist->nri + nri_max > nlist->maxnri)
+ {
+ nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
+ reallocate_nblist(nlist);
+ }
+
+ /* Loop over the atoms in the i super-cluster */
+ for (c = 0; c < GPU_NSUBCELL; c++)
+ {
+ c_abs = sci*GPU_NSUBCELL + c;
+
+ for (i = 0; i < nbl->na_ci; i++)
+ {
+ ind_i = c_abs*nbl->na_ci + i;
+ ai = nbs->a[ind_i];
+ if (ai >= 0)
+ {
+ nri = nlist->nri;
+ nlist->jindex[nri+1] = nlist->jindex[nri];
+ nlist->iinr[nri] = ai;
+ /* With GPUs, energy groups are not supported */
+ nlist->gid[nri] = 0;
+ nlist->shift[nri] = nbl_sci->shift & NBNXN_CI_SHIFT;
+
+ bFEP_i = (gridi->fep[c_abs - gridi->cell0] & (1 << i));
+
+ xi = nbat->x[ind_i*nbat->xstride+XX] + shx;
+ yi = nbat->x[ind_i*nbat->xstride+YY] + shy;
+ zi = nbat->x[ind_i*nbat->xstride+ZZ] + shz;
+
+ if ((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE*nbl->na_cj > nlist->maxnrj)
+ {
+ nlist->maxnrj = over_alloc_small((nlist->nrj + cj4_ind_end - cj4_ind_start)*NBNXN_GPU_JGROUP_SIZE*nbl->na_cj);
+ srenew(nlist->jjnr, nlist->maxnrj);
+ srenew(nlist->excl_fep, nlist->maxnrj);
+ }
+
+ for (cj4_ind = cj4_ind_start; cj4_ind < cj4_ind_end; cj4_ind++)
+ {
+ cj4 = &nbl->cj4[cj4_ind];
+
+ for (gcj = 0; gcj < NBNXN_GPU_JGROUP_SIZE; gcj++)
+ {
+ unsigned int fep_cj;
+
+ if ((cj4->imei[0].imask & (1U << (gcj*GPU_NSUBCELL + c))) == 0)
+ {
+ /* Skip this ci for this cj */
+ continue;
+ }
+
+ cjr = cj4->cj[gcj] - gridj->cell0*GPU_NSUBCELL;
+
+ fep_cj = gridj->fep[cjr];
+
+ if (bFEP_i || fep_cj != 0)
+ {
+ for (j = 0; j < nbl->na_cj; j++)
+ {
+ /* Is this interaction perturbed and not excluded? */
+ ind_j = (gridj->cell0*GPU_NSUBCELL + cjr)*nbl->na_cj + j;
+ aj = nbs->a[ind_j];
+ if (aj >= 0 &&
+ (bFEP_i || (fep_cj & (1 << j))) &&
+ (!bDiagRemoved || ind_j >= ind_i))
+ {
+ nbnxn_excl_t *excl;
+ int excl_pair;
+ unsigned int excl_bit;
+ real dx, dy, dz;
+
+ get_nbl_exclusions_1(nbl, cj4_ind, j>>2, &excl);
+
+ excl_pair = a_mod_wj(j)*nbl->na_ci + i;
+ excl_bit = (1U << (gcj*GPU_NSUBCELL + c));
+
+ dx = nbat->x[ind_j*nbat->xstride+XX] - xi;
+ dy = nbat->x[ind_j*nbat->xstride+YY] - yi;
+ dz = nbat->x[ind_j*nbat->xstride+ZZ] - zi;
+
+ /* The unpruned GPU list has more than 2/3
+ * of the atom pairs beyond rlist. Using
+ * this list will cause a lot of overhead
+ * in the CPU FEP kernels, especially
+ * relative to the fast GPU kernels.
+ * So we prune the FEP list here.
+ */
+ if (dx*dx + dy*dy + dz*dz < rlist_fep2)
+ {
+ if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
+ {
+ fep_list_new_nri_copy(nlist);
+ nri = nlist->nri;
+ }
+
+ /* Add it to the FEP list */
+ nlist->jjnr[nlist->nrj] = aj;
+ nlist->excl_fep[nlist->nrj] = (excl->pair[excl_pair] & excl_bit) ? 1 : 0;
+ nlist->nrj++;
+ }
+
+ /* Exclude it from the normal list.
+ * Note that the charge and LJ parameters have
+ * been set to zero, but we need to avoid 0/0,
+ * as perturbed atoms can be on top of each other.
+ */
+ excl->pair[excl_pair] &= ~excl_bit;
+ }
+ }
+
+ /* Note that we could mask out this pair in imask
+ * if all i- and/or all j-particles are perturbed.
+ * But since the perturbed pairs on the CPU will
+ * take an order of magnitude more time, the GPU
+ * will finish before the CPU and there is no gain.
+ */
+ }
+ }
+ }
+
+ if (nlist->nrj > nlist->jindex[nri])
+ {
+ /* Actually add this new, non-empty, list */
+ nlist->nri++;
+ nlist->jindex[nlist->nri] = nlist->nrj;
+ }
+ }
+ }
+ }
+}
+
/* Set all atom-pair exclusions from the topology stored in excl
* as masks in the pair-list for i-super-cell entry nbl_sci
*/
inner_i = i - si*na_c;
inner_e = ge - se*na_c;
-/* Macro for getting the index of atom a within a cluster */
-#define AMODCJ4(a) ((a) & (NBNXN_GPU_JGROUP_SIZE - 1))
-/* Macro for converting an atom number to a cluster number */
-#define A2CJ4(a) ((a) >> NBNXN_GPU_JGROUP_SIZE_2LOG)
-/* Macro for getting the index of an i-atom within a warp */
-#define AMODWI(a) ((a) & (NBNXN_GPU_CLUSTER_SIZE/2 - 1))
-
- if (nbl_imask0(nbl, found) & (1U << (AMODCJ4(found)*GPU_NSUBCELL + si)))
+ if (nbl_imask0(nbl, found) & (1U << (cj_mod_cj4(found)*GPU_NSUBCELL + si)))
{
w = (inner_e >> 2);
- get_nbl_exclusions_1(nbl, A2CJ4(found), w, &nbl_excl);
+ get_nbl_exclusions_1(nbl, cj_to_cj4(found), w, &nbl_excl);
- nbl_excl->pair[AMODWI(inner_e)*nbl->na_ci+inner_i] &=
- ~(1U << (AMODCJ4(found)*GPU_NSUBCELL + si));
+ nbl_excl->pair[a_mod_wj(inner_e)*nbl->na_ci+inner_i] &=
+ ~(1U << (cj_mod_cj4(found)*GPU_NSUBCELL + si));
}
-
-#undef AMODCJ4
-#undef A2CJ4
-#undef AMODWI
}
}
}
nbl->work->ncj_hlj = 0;
}
+/* Clears a group scheme pair list */
+static void clear_pairlist_fep(t_nblist *nl)
+{
+ nl->nri = 0;
+ nl->nrj = 0;
+ if (nl->jindex == NULL)
+ {
+ snew(nl->jindex, 1);
+ }
+ nl->jindex[0] = 0;
+}
+
/* Sets a simple list i-cell bounding box, including PBC shift */
static gmx_inline void set_icell_bb_simple(const nbnxn_bb_t *bb, int ci,
real shx, real shy, real shz,
int stride, const real *x,
nbnxn_list_work_t *work)
{
- int ia, i;
+ int ia, i;
real *x_ci;
x_ci = work->x_ci;
int stride, const real *x,
nbnxn_list_work_t *work)
{
- int si, io, ia, i, j;
+ int si, io, ia, i, j;
real *x_ci;
x_ci = work->x_ci;
}
#endif
+static real minimum_subgrid_size_xy(const nbnxn_grid_t *grid)
+{
+ if (grid->bSimple)
+ {
+ return min(grid->sx, grid->sy);
+ }
+ else
+ {
+ return min(grid->sx/GPU_NSUBCELL_X, grid->sy/GPU_NSUBCELL_Y);
+ }
+}
+
+static real effective_buffer_1x1_vs_MxN(const nbnxn_grid_t *gridi,
+ const nbnxn_grid_t *gridj)
+{
+ const real eff_1x1_buffer_fac_overest = 0.1;
+
+ /* Determine an atom-pair list cut-off buffer size for atom pairs,
+ * to be added to rlist (including buffer) used for MxN.
+ * This is for converting an MxN list to a 1x1 list. This means we can't
+ * use the normal buffer estimate, as we have an MxN list in which
+ * some atom pairs beyond rlist are missing. We want to capture
+ * the beneficial effect of buffering by extra pairs just outside rlist,
+ * while removing the useless pairs that are further away from rlist.
+ * (Also the buffer could have been set manually not using the estimate.)
+ * This buffer size is an overestimate.
+ * We add 10% of the smallest grid sub-cell dimensions.
+ * Note that the z-size differs per cell and we don't use this,
+ * so we overestimate.
+ * With PME, the 10% value gives a buffer that is somewhat larger
+ * than the effective buffer with a tolerance of 0.005 kJ/mol/ps.
+ * Smaller tolerances or using RF lead to a smaller effective buffer,
+ * so 10% gives a safe overestimate.
+ */
+ return eff_1x1_buffer_fac_overest*(minimum_subgrid_size_xy(gridi) +
+ minimum_subgrid_size_xy(gridj));
+}
+
/* Clusters at the cut-off only increase rlist by 60% of their size */
static real nbnxn_rlist_inc_outside_fac = 0.6;
int min_ci_balanced)
{
const nbnxn_grid_t *grid;
- rvec ls;
- real xy_diag2, r_eff_sup, vol_est, nsp_est, nsp_est_nl;
- int nsubpair_max;
+ rvec ls;
+ real xy_diag2, r_eff_sup, vol_est, nsp_est, nsp_est_nl;
+ int nsubpair_max;
grid = &nbs->grid[0];
{
int nsci, ncj4, nexcl;
int n, i;
+ int nthreads gmx_unused;
if (nblc->bSimple)
{
/* Each thread should copy its own data to the combined arrays,
* as otherwise data will go back and forth between different caches.
*/
-#pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
+ nthreads = gmx_omp_nthreads_get(emntPairsearch);
+#pragma omp parallel for num_threads(nthreads) schedule(static)
for (n = 0; n < nnbl; n++)
{
- int sci_offset;
- int cj4_offset;
- int ci_offset;
- int excl_offset;
- int i, j4;
+ int sci_offset;
+ int cj4_offset;
+ int ci_offset;
+ int excl_offset;
+ int i, j4;
const nbnxn_pairlist_t *nbli;
/* Determine the offset in the combined data for our thread */
}
}
+static void balance_fep_lists(const nbnxn_search_t nbs,
+ nbnxn_pairlist_set_t *nbl_lists)
+{
+ int nnbl, th;
+ int nri_tot, nrj_tot, nrj_target;
+ int th_dest;
+ t_nblist *nbld;
+
+ nnbl = nbl_lists->nnbl;
+
+ if (nnbl == 1)
+ {
+ /* Nothing to balance */
+ return;
+ }
+
+ /* Count the total i-lists and pairs */
+ nri_tot = 0;
+ nrj_tot = 0;
+ for (th = 0; th < nnbl; th++)
+ {
+ nri_tot += nbl_lists->nbl_fep[th]->nri;
+ nrj_tot += nbl_lists->nbl_fep[th]->nrj;
+ }
+
+ nrj_target = (nrj_tot + nnbl - 1)/nnbl;
+
+ assert(gmx_omp_nthreads_get(emntNonbonded) == nnbl);
+
+#pragma omp parallel for schedule(static) num_threads(nnbl)
+ for (th = 0; th < nnbl; th++)
+ {
+ t_nblist *nbl;
+
+ nbl = nbs->work[th].nbl_fep;
+
+ /* Note that here we allocate for the total size, instead of
+ * a per-thread esimate (which is hard to obtain).
+ */
+ if (nri_tot > nbl->maxnri)
+ {
+ nbl->maxnri = over_alloc_large(nri_tot);
+ reallocate_nblist(nbl);
+ }
+ if (nri_tot > nbl->maxnri || nrj_tot > nbl->maxnrj)
+ {
+ nbl->maxnrj = over_alloc_small(nrj_tot);
+ srenew(nbl->jjnr, nbl->maxnrj);
+ srenew(nbl->excl_fep, nbl->maxnrj);
+ }
+
+ clear_pairlist_fep(nbl);
+ }
+
+ /* Loop over the source lists and assign and copy i-entries */
+ th_dest = 0;
+ nbld = nbs->work[th_dest].nbl_fep;
+ for (th = 0; th < nnbl; th++)
+ {
+ t_nblist *nbls;
+ int i, j;
+
+ nbls = nbl_lists->nbl_fep[th];
+
+ for (i = 0; i < nbls->nri; i++)
+ {
+ int nrj;
+
+ /* The number of pairs in this i-entry */
+ nrj = nbls->jindex[i+1] - nbls->jindex[i];
+
+ /* Decide if list th_dest is too large and we should procede
+ * to the next destination list.
+ */
+ if (th_dest+1 < nnbl && nbld->nrj > 0 &&
+ nbld->nrj + nrj - nrj_target > nrj_target - nbld->nrj)
+ {
+ th_dest++;
+ nbld = nbs->work[th_dest].nbl_fep;
+ }
+
+ nbld->iinr[nbld->nri] = nbls->iinr[i];
+ nbld->gid[nbld->nri] = nbls->gid[i];
+ nbld->shift[nbld->nri] = nbls->shift[i];
+
+ for (j = nbls->jindex[i]; j < nbls->jindex[i+1]; j++)
+ {
+ nbld->jjnr[nbld->nrj] = nbls->jjnr[j];
+ nbld->excl_fep[nbld->nrj] = nbls->excl_fep[j];
+ nbld->nrj++;
+ }
+ nbld->nri++;
+ nbld->jindex[nbld->nri] = nbld->nrj;
+ }
+ }
+
+ /* Swap the list pointers */
+ for (th = 0; th < nnbl; th++)
+ {
+ t_nblist *nbl_tmp;
+
+ nbl_tmp = nbl_lists->nbl_fep[th];
+ nbl_lists->nbl_fep[th] = nbs->work[th].nbl_fep;
+ nbs->work[th].nbl_fep = nbl_tmp;
+
+ if (debug)
+ {
+ fprintf(debug, "nbl_fep[%d] nri %4d nrj %4d\n",
+ th,
+ nbl_lists->nbl_fep[th]->nri,
+ nbl_lists->nbl_fep[th]->nrj);
+ }
+ }
+}
+
/* Returns the next ci to be processes by our thread */
static gmx_bool next_ci(const nbnxn_grid_t *grid,
int conv,
const int ci_block_enum = 5;
const int ci_block_denom = 11;
const int ci_block_min_atoms = 16;
- int ci_block;
+ int ci_block;
/* Here we decide how to distribute the blocks over the threads.
* We use prime numbers to try to avoid that the grid size becomes
gmx_bool progBal,
int min_ci_balanced,
int th, int nth,
- nbnxn_pairlist_t *nbl)
+ nbnxn_pairlist_t *nbl,
+ t_nblist *nbl_fep)
{
- int na_cj_2log;
- matrix box;
- real rl2;
- float rbb2;
- int d;
- int ci_b, ci, ci_x, ci_y, ci_xy, cj;
- ivec shp;
- int tx, ty, tz;
- int shift;
- gmx_bool bMakeList;
- real shx, shy, shz;
- int conv_i, cell0_i;
+ int na_cj_2log;
+ matrix box;
+ real rl2, rl_fep2 = 0;
+ float rbb2;
+ int d;
+ int ci_b, ci, ci_x, ci_y, ci_xy, cj;
+ ivec shp;
+ int tx, ty, tz;
+ int shift;
+ gmx_bool bMakeList;
+ real shx, shy, shz;
+ int conv_i, cell0_i;
const nbnxn_bb_t *bb_i = NULL;
#ifdef NBNXN_BBXXXX
- const float *pbb_i = NULL;
+ const float *pbb_i = NULL;
#endif
- const float *bbcz_i, *bbcz_j;
- const int *flags_i;
- real bx0, bx1, by0, by1, bz0, bz1;
- real bz1_frac;
- real d2cx, d2z, d2z_cx, d2z_cy, d2zx, d2zxy, d2xy;
- int cxf, cxl, cyf, cyf_x, cyl;
- int cx, cy;
- int c0, c1, cs, cf, cl;
- int ndistc;
- int ncpcheck;
- int gridi_flag_shift = 0, gridj_flag_shift = 0;
- unsigned *gridj_flag = NULL;
- int ncj_old_i, ncj_old_j;
+ const float *bbcz_i, *bbcz_j;
+ const int *flags_i;
+ real bx0, bx1, by0, by1, bz0, bz1;
+ real bz1_frac;
+ real d2cx, d2z, d2z_cx, d2z_cy, d2zx, d2zxy, d2xy;
+ int cxf, cxl, cyf, cyf_x, cyl;
+ int cx, cy;
+ int c0, c1, cs, cf, cl;
+ int ndistc;
+ int ncpcheck;
+ int gridi_flag_shift = 0, gridj_flag_shift = 0;
+ unsigned int *gridj_flag = NULL;
+ int ncj_old_i, ncj_old_j;
nbs_cycle_start(&work->cc[enbsCCsearch]);
rl2 = nbl->rlist*nbl->rlist;
+ if (nbs->bFEP && !nbl->bSimple)
+ {
+ /* Determine an atom-pair list cut-off distance for FEP atom pairs.
+ * We should not simply use rlist, since then we would not have
+ * the small, effective buffering of the NxN lists.
+ * The buffer is on overestimate, but the resulting cost for pairs
+ * beyond rlist is neglible compared to the FEP pairs within rlist.
+ */
+ rl_fep2 = nbl->rlist + effective_buffer_1x1_vs_MxN(gridi, gridj);
+
+ if (debug)
+ {
+ fprintf(debug, "nbl_fep atom-pair rlist %f\n", rl_fep2);
+ }
+ rl_fep2 = rl_fep2*rl_fep2;
+ }
+
rbb2 = boundingbox_only_distance2(gridi, gridj, nbl->rlist, nbl->bSimple);
if (debug)
na_cj_2log,
&(nbl->ci[nbl->nci]),
excl);
+
+ if (nbs->bFEP)
+ {
+ make_fep_list(nbs, nbat, nbl,
+ shift == CENTRAL && gridi == gridj,
+ &(nbl->ci[nbl->nci]),
+ gridi, gridj, nbl_fep);
+ }
}
else
{
gridj->na_c_2log,
&(nbl->sci[nbl->nsci]),
excl);
+
+ if (nbs->bFEP)
+ {
+ make_fep_list_supersub(nbs, nbat, nbl,
+ shift == CENTRAL && gridi == gridj,
+ &(nbl->sci[nbl->nsci]),
+ shx, shy, shz,
+ rl_fep2,
+ gridi, gridj, nbl_fep);
+ }
}
/* Close this ci list */
print_nblist_statistics_supersub(debug, nbl, nbs, rlist);
}
+ if (nbs->bFEP)
+ {
+ fprintf(debug, "nbl FEP list pairs: %d\n", nbl_fep->nrj);
+ }
}
}
int nsrc,
const nbnxn_buffer_flags_t *dest)
{
- int s, b;
- const unsigned *flag;
+ int s, b;
+ const unsigned int *flag;
for (s = 0; s < nsrc; s++)
{
int nb_kernel_type,
t_nrnb *nrnb)
{
- nbnxn_grid_t *gridi, *gridj;
- gmx_bool bGPUCPU;
- int nzi, zi, zj0, zj1, zj;
- int nsubpair_max;
- int th;
- int nnbl;
+ nbnxn_grid_t *gridi, *gridj;
+ gmx_bool bGPUCPU;
+ int nzi, zi, zj0, zj1, zj;
+ int nsubpair_max;
+ int th;
+ int nnbl;
nbnxn_pairlist_t **nbl;
- int ci_block;
- gmx_bool CombineNBLists;
- gmx_bool progBal;
- int np_tot, np_noq, np_hlj, nap;
+ int ci_block;
+ gmx_bool CombineNBLists;
+ gmx_bool progBal;
+ int np_tot, np_noq, np_hlj, nap;
/* Check if we are running hybrid GPU + CPU nbnxn mode */
bGPUCPU = (!nbs->grid[0].bSimple && nbl_list->bSimple);
for (th = 0; th < nnbl; th++)
{
clear_pairlist(nbl[th]);
+
+ if (nbs->bFEP)
+ {
+ clear_pairlist_fep(nbl_list->nbl_fep[th]);
+ }
}
for (zi = 0; zi < nzi; zi++)
nsubpair_max,
progBal, min_ci_balanced,
th, nnbl,
- nbl[th]);
+ nbl[th],
+ nbl_list->nbl_fep[th]);
}
nbs_cycle_stop(&nbs->cc[enbsCCsearch]);
reduce_buffer_flags(nbs, nnbl, &nbat->buffer_flags);
}
+ if (nbs->bFEP)
+ {
+ /* Balance the free-energy lists over all the threads */
+ balance_fep_lists(nbs, nbl_list);
+ }
+
/* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
if (LOCAL_I(iloc))
{