::std::vector<real>
simd4Real2Vector(const gmx_simd4_real_t simd4)
{
real mem[GMX_SIMD4_WIDTH*2];
real * p = gmx_simd4_align_r(mem);
gmx_simd4_store_r(p, simd4);
std::vector<real> v(p, p+GMX_SIMD4_WIDTH);
return v;
}
gmx_simd4_real_t
vector2Simd4Real(const std::vector<real> &v)
{
real mem[GMX_SIMD4_WIDTH*2];
real * p = gmx_simd4_align_r(mem);
for (int i = 0; i < GMX_SIMD4_WIDTH; i++)
{
p[i] = v[i % v.size()]; // repeat vector contents to fill simd width
}
return gmx_simd4_load_r(p);
}
struct pme_spline_work *make_pme_spline_work(int gmx_unused order)
{
struct pme_spline_work *work;
#ifdef PME_SIMD4_SPREAD_GATHER
real tmp[GMX_SIMD4_WIDTH*3], *tmp_aligned;
gmx_simd4_real_t zero_S;
gmx_simd4_real_t real_mask_S0, real_mask_S1;
int of, i;
snew_aligned(work, 1, SIMD4_ALIGNMENT);
tmp_aligned = gmx_simd4_align_r(tmp);
zero_S = gmx_simd4_setzero_r();
/* Generate bit masks to mask out the unused grid entries,
* as we only operate on order of the 8 grid entries that are
* load into 2 SIMD registers.
*/
for (of = 0; of < 2*GMX_SIMD4_WIDTH-(order-1); of++)
{
for (i = 0; i < 2*GMX_SIMD4_WIDTH; i++)
{
tmp_aligned[i] = (i >= of && i < of+order ? -1.0 : 1.0);
}
real_mask_S0 = gmx_simd4_load_r(tmp_aligned);
real_mask_S1 = gmx_simd4_load_r(tmp_aligned+GMX_SIMD4_WIDTH);
work->mask_S0[of] = gmx_simd4_cmplt_r(real_mask_S0, zero_S);
work->mask_S1[of] = gmx_simd4_cmplt_r(real_mask_S1, zero_S);
}
#else
work = NULL;
#endif
return work;
}
void gather_f_bsplines(struct gmx_pme_t *pme, real *grid,
gmx_bool bClearF, pme_atomcomm_t *atc,
splinedata_t *spline,
real scale)
{
/* sum forces for local particles */
int nn, n, ithx, ithy, ithz, i0, j0, k0;
int index_x, index_xy;
int nx, ny, nz, pny, pnz;
int * idxptr;
real tx, ty, dx, dy, coefficient;
real fx, fy, fz, gval;
real fxy1, fz1;
real *thx, *thy, *thz, *dthx, *dthy, *dthz;
int norder;
real rxx, ryx, ryy, rzx, rzy, rzz;
int order;
#ifdef PME_SIMD4_SPREAD_GATHER
// cppcheck-suppress unreadVariable cppcheck seems not to analyze code from pme-simd4.h
struct pme_spline_work *work = pme->spline_work;
#ifndef PME_SIMD4_UNALIGNED
real thz_buffer[GMX_SIMD4_WIDTH*3], *thz_aligned;
real dthz_buffer[GMX_SIMD4_WIDTH*3], *dthz_aligned;
thz_aligned = gmx_simd4_align_r(thz_buffer);
dthz_aligned = gmx_simd4_align_r(dthz_buffer);
#endif
#endif
order = pme->pme_order;
nx = pme->nkx;
ny = pme->nky;
nz = pme->nkz;
pny = pme->pmegrid_ny;
pnz = pme->pmegrid_nz;
rxx = pme->recipbox[XX][XX];
ryx = pme->recipbox[YY][XX];
ryy = pme->recipbox[YY][YY];
rzx = pme->recipbox[ZZ][XX];
rzy = pme->recipbox[ZZ][YY];
rzz = pme->recipbox[ZZ][ZZ];
for (nn = 0; nn < spline->n; nn++)
{
n = spline->ind[nn];
coefficient = scale*atc->coefficient[n];
if (bClearF)
{
atc->f[n][XX] = 0;
atc->f[n][YY] = 0;
atc->f[n][ZZ] = 0;
}
if (coefficient != 0)
{
fx = 0;
fy = 0;
fz = 0;
idxptr = atc->idx[n];
norder = nn*order;
i0 = idxptr[XX];
j0 = idxptr[YY];
k0 = idxptr[ZZ];
/* Pointer arithmetic alert, next six statements */
thx = spline->theta[XX] + norder;
thy = spline->theta[YY] + norder;
thz = spline->theta[ZZ] + norder;
dthx = spline->dtheta[XX] + norder;
dthy = spline->dtheta[YY] + norder;
dthz = spline->dtheta[ZZ] + norder;
switch (order)
{
case 4:
#ifdef PME_SIMD4_SPREAD_GATHER
#ifdef PME_SIMD4_UNALIGNED
#define PME_GATHER_F_SIMD4_ORDER4
#else
#define PME_GATHER_F_SIMD4_ALIGNED
#define PME_ORDER 4
#endif
#include "gromacs/ewald/pme-simd4.h"
#else
DO_FSPLINE(4);
#endif
break;
case 5:
#ifdef PME_SIMD4_SPREAD_GATHER
#define PME_GATHER_F_SIMD4_ALIGNED
#define PME_ORDER 5
#include "gromacs/ewald/pme-simd4.h"
#else
DO_FSPLINE(5);
#endif
break;
default:
DO_FSPLINE(order);
break;
}
atc->f[n][XX] += -coefficient*( fx*nx*rxx );
atc->f[n][YY] += -coefficient*( fx*nx*ryx + fy*ny*ryy );
atc->f[n][ZZ] += -coefficient*( fx*nx*rzx + fy*ny*rzy + fz*nz*rzz );
}
}
/* Since the energy and not forces are interpolated
* the net force might not be exactly zero.
* This can be solved by also interpolating F, but
* that comes at a cost.
* A better hack is to remove the net force every
* step, but that must be done at a higher level
* since this routine doesn't see all atoms if running
* in parallel. Don't know how important it is? EL 990726
*/
}
